An advertisement delivery strategy optimization method, system and platform

Abstract

The application provides an advertisement putting strategy optimization method, system and platform, and belongs to the technical field of advertisement putting, which comprises the following steps: obtaining multi-dimensional attribute information of each advertisement position on a target display screen from a record database, performing quantitative processing, and constructing an attribute representation matrix; extracting a representation vector of each attribute representation matrix, and obtaining multi-dimensional features of each advertisement position in dependence on original distribution of the advertisement positions on the target display screen; matching a historical change baseline matched with each multi-dimensional feature from a historical putting database, and predicting putting advantages of the corresponding advertisement positions in a preset time period in dependence on the historical change baseline and in combination with a virtual advertisement; performing time alignment processing according to the putting advantages of all the advertisement positions, determining advertisement putting highlight coefficients of each alignment time point, locking time points with highlight coefficients greater than a preset coefficient as cutting points, and optimizing the advertisement putting strategy. The application effectively improves the pertinence and effect of advertisement putting, and reduces resource waste.

Description

Technical Field

[0001] This invention relates to the field of advertising delivery technology, and in particular to a method, system and platform for optimizing advertising delivery strategies. Background Technology

[0002] Current advertising placement technologies suffer from several limitations in ad placement value assessment and strategy optimization: Analysis of ad placement attributes is often limited to a single dimension (e.g., focusing only on play counts or impressions), resulting in a one-sided characterization of ad placement features and failing to reflect true placement potential. Furthermore, existing methods often evaluate individual ad placements in isolation, neglecting the spatial relationships and mutual influences between ad placements on the same screen (e.g., visual interference or synergistic effects between adjacent ad placements). This leads to a single feature dimension, failing to accurately capture the dynamic value of ad placements. Traditional methods also rely heavily on direct transfer of historical data, lacking a dynamic correlation mechanism between historical data and the real-time placement environment, resulting in significant discrepancies between predicted results and actual placement effectiveness. During playback, reliance on empirical time segmentation leads to a mismatch between ad placement and peak user attention, resulting in low resource utilization. Moreover, insufficient handling of interfering factors such as fraudulent orders further impacts strategy effectiveness, leading to low advertising placement efficiency.

[0003] Therefore, this invention proposes a method, system, and platform for optimizing advertising delivery strategies. Summary of the Invention

[0004] This invention provides a method, system, and platform for optimizing advertising placement strategies, which integrates multi-dimensional attributes, mines the correlation features of ad placements, combines virtual advertising with historical data to accurately predict placement advantages, and optimizes placement effectiveness through time nodes.

[0005] This invention provides a method for optimizing advertising delivery strategies, comprising:

[0006] Step 1: Obtain multi-dimensional attribute information of each ad slot on the target display screen from the record database and perform quantitative processing to construct an attribute representation matrix. The multi-dimensional attribute information is related to the pre-ad active playback information, the pre-ad passive playback information, the traffic volume based on each passive playback, the added virtual ads, and the playback status of the virtual ads.

[0007] Step 2: Extract the representation vector of each attribute representation matrix, and based on the original distribution of the ad slots on the target display screen, obtain the multidimensional features of each ad slot, wherein the multidimensional features include: main representation features and auxiliary representation features of the other ad slots for the corresponding ad slots;

[0008] Step 3: Match the historical change baseline that matches each multi-dimensional feature from the historical delivery database, and rely on the historical change baseline and combine it with the added virtual ads to predict the delivery advantage of the corresponding ad slot within the preset time period;

[0009] Step 4: Perform time alignment based on the advantages of all ad placements, determine the ad placement prominence coefficient for each aligned time point, and lock the time points where the prominence coefficient is greater than the preset coefficient as the cut-off points to optimize the ad placement strategy.

[0010] Preferably, the acquired virtual advertisements include:

[0011] Acquire the generated creatives, visual effects, and placement environment for each ad placement based on the target display screen, and construct the placement bias for each ad placement;

[0012] Determine the first sequence of historical order-brushing attributes and the second sequence of targeting bias for each ad placement;

[0013] Extract the feature parameters of the first sequence and the second sequence, determine the connection relationship between the feature parameters, and add the initial advertisement to the corresponding advertisement position;

[0014] The initial advertisement is split into single frames, a directed graph of order-brushing elimination is constructed for each split frame, and the elimination probability is obtained;

[0015] The initial advertisement is divided using an interactive window, and the elimination characteristics of each sub-advertisement are determined by combining the elimination probability.

[0016] Linear detection is performed on all eliminated features to extract the best features. The best features are combined with the global coverage ratio and trapping strength to optimize the initial advertisement and obtain the virtual advertisement.

[0017] Preferably, the multi-dimensional attribute information of each advertisement slot on the target display screen is obtained from the record database and quantified to construct an attribute representation matrix, including:

[0018] The multidimensional attribute information is input into a quantization processing model that matches the target display screen to obtain the multidimensional output vector of the corresponding ad position. At the same time, the user behavior dynamic features of each ad position and the spatial correlation parameters of adjacent ad positions are obtained as extended dimensions of the multidimensional attribute information and quantized to obtain the extended vector.

[0019] Based on the mapping relationship between each extended dimension and each dimension type in the multidimensional attribute information, the mapping priority is determined, and a mapping path is generated;

[0020] Based on the mapping path and the set of ad variations for each dimension type, a topology network is constructed.

[0021] The topology network is analyzed to determine the parameters to be retained in the extended dimension and the influence factors of each retained parameter on the corresponding original elements under the mapping relationship. The new vector is then obtained.

[0022] Map all new vectors to a high-dimensional feature space to generate attribute representation matrices.

[0023] Preferably, based on the original distribution of ad slots on the target display screen, multidimensional features of each ad slot are obtained, including:

[0024] The attribute representation matrix is ​​subjected to nonlinear dimensionality reduction to extract a deep principal representation vector containing high-order feature interactions;

[0025] A spatial association graph of ad placements is constructed based on the physical coordinate distribution of the target display screen, where nodes are ad placements and edge weights are the visual attention shift probability and spatiotemporal co-occurrence frequency of adjacent ad placements.

[0026] Graph neural networks are used to propagate features in a spatial relational graph. Noise features of nodes and path segments between nodes are extracted during the propagation process to construct a noise sequence and extract local and global significant noise factors.

[0027] The feature propagation process is noise-eliminating based on all local and global significant noise factors, generating spatial correlation feature vectors for each ad slot as auxiliary representation features. In the feature propagation process, a distance attenuation coefficient is introduced so that the feature influence of adjacent ad slots decreases linearly with the increase of physical distance.

[0028] By combining the characteristic fluctuation sequence of the ad placement within a preset historical period, a time-series dynamic feature vector is extracted as a supplementary auxiliary representation feature.

[0029] Based on deep master representation vectors, spatial correlation feature vectors, and temporal dynamic feature vectors, multidimensional features of corresponding ad slots are generated.

[0030] Preferably, relying on the historical change baseline and combining the added virtual advertising to predict the placement advantage of the corresponding ad slot within a preset time period, includes:

[0031] The historical change baseline is decomposed into multiple sub-baseline sequences according to the time granularity. Each sub-baseline sequence corresponds to a time unit within a preset time period. The trend characteristic parameters and fluctuation thresholds of each sub-baseline sequence are extracted.

[0032] The playback data of the newly added virtual advertisements is analyzed in layers to obtain the interaction conversion rate and content retention rate of virtual advertisements in different user profile groups, and a two-dimensional profile matrix of virtual advertisements is constructed.

[0033] The elements in each dimension of the two-dimensional profile matrix are sorted in descending order according to their values, and a sequence number group is constructed for each remaining user except for the last user in the sort.

[0034] A difference matrix is ​​constructed based on the difference between two adjacent rows of the matrix after descending order. The results of each column in the difference matrix are clustered and classified, and the mode range of each column is determined. The minimum value in the mode range is used to replace the element values ​​before the mode range, and the maximum value in the mode range is used to replace the element values ​​after the mode range to obtain a new two-dimensional matrix. The two-dimensional arrays of the remaining users are then obtained.

[0035] The effect mapping vector is obtained by relying on the two-dimensional array, the index group and the original array of each remaining user;

[0036] The fusion weights of the historical change baseline and the effect mapping vector are dynamically adjusted based on the time decay function. The weight of the effect mapping vector increases exponentially with the increase of real-time interactive data, while the weight of the historical change baseline decreases linearly with the extension of the time interval.

[0037] The fused feature vector and environmental variable parameters for a preset time period are input into the delivery advantage prediction model, and the delivery advantage quantification value for each time unit is output.

[0038] The quantitative value of the delivery advantage in continuous time units is smoothed to generate a delivery advantage curve that includes local peaks and global trends. The area under the curve and the duration of the peak are used as the final evaluation indicators of the delivery advantage.

[0039] Preferably, the ad highlighting factor for each aligned time point is determined, including:

[0040] Based on the quantitative value of the placement advantage of the time unit to which the alignment time point belongs, and combined with the placement interference coefficient of other ad positions at the corresponding time point, a basic prominence coefficient is constructed.

[0041] Introduce user activity weights corresponding to specific time points to dynamically adjust the basic prominence coefficient;

[0042] Extract the core visual features and content timeliness parameters of the ad placement at the corresponding time point, and generate feature enhancement factors;

[0043] The corrected salience coefficient and the feature enhancement factor are nonlinearly fused together. At the same time, the simulated interactive feedback of virtual ads at the corresponding time points is combined for secondary calibration to obtain the final ad placement salience coefficient.

[0044] Specifically, when the interference coefficient of an adjacent ad slot exceeds a first preset threshold, the feature enhancement factor weight of the current ad slot is automatically increased.

[0045] Preferably, the advertising strategy is optimized, including:

[0046] Based on the locked cut points, several advertising time segments are divided. Combining the advantages of advertising placements, user profile tags, and simulated conversion data of virtual ads within each time segment, a value assessment matrix for each segment is constructed.

[0047] Based on the value assessment matrix, the ad slots in each time segment are prioritized and the ad playback duration and frequency are dynamically adjusted.

[0048] Real-time public opinion data and competitor advertising dynamics are introduced as feedback factors. When the feedback factor triggers the second preset threshold, the backup advertising material library is automatically activated to replace advertising content with low conversion potential.

[0049] By comparing the ad exposure conversion rate, the increase in user dwell time, and the performance deviation rate between virtual and real ads before and after optimization, the threshold for determining the cut-off point and the granularity of time segment division are iteratively adjusted to form a self-optimizing ad placement strategy.

[0050] This invention provides an optimization system for advertising delivery strategies, comprising:

[0051] The matrix construction module is used to obtain multi-dimensional attribute information of each ad slot on the target display screen from the record database and perform quantitative processing to construct an attribute representation matrix. The multi-dimensional attribute information is related to pre-ad active playback information, pre-ad passive playback information, traffic volume based on each passive playback, added virtual ads, and the playback status of virtual ads.

[0052] The feature acquisition module is used to extract the representation vector of each attribute representation matrix and, depending on the original distribution of the ad slots on the target display screen, obtain the multidimensional features of each ad slot. The multidimensional features include: main representation features and auxiliary representation features of the remaining ad slots for the corresponding ad slots.

[0053] The advantage determination module is used to match historical change baselines that match each multi-dimensional feature from the historical delivery database, and based on the historical change baselines, combined with the added virtual advertisements, predict the delivery advantage of the corresponding advertisement position within a preset time period.

[0054] The segmentation optimization module is used to perform time alignment processing based on the advantages of all ad placements, determine the ad placement prominence coefficient at each aligned time point, and lock the time points with prominence coefficients greater than preset coefficients as segmentation points to optimize the ad placement strategy.

[0055] This invention provides a platform for executing an optimization method for any of the described advertising delivery strategies.

[0056] Compared with the prior art, the beneficial effects of this application are as follows:

[0057] By comprehensively quantifying ad placement characteristics using multi-dimensional attributes, mining the correlation between ad placements, integrating historical data with the advantages of virtual advertising prediction, and precisely optimizing based on time nodes, the targeting and effectiveness of advertising are effectively improved, resource waste is reduced, and the strategy's anti-interference capability and adaptability are enhanced through virtual advertising and interference handling.

[0058] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0059] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of an optimization method for an advertising delivery strategy according to an embodiment of the present invention;

[0062] Figure 2 This is a structural diagram of an advertising delivery strategy optimization system according to an embodiment of the present invention. Detailed Implementation

[0063] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0064] This invention provides a method for optimizing advertising delivery strategies, such as... Figure 1 As shown, it includes:

[0065] Step 1: Obtain multi-dimensional attribute information of each ad slot on the target display screen from the record database and perform quantitative processing to construct an attribute representation matrix. The multi-dimensional attribute information is related to the pre-ad active playback information, the pre-ad passive playback information, the traffic volume based on each passive playback, the added virtual ads, and the playback status of the virtual ads.

[0066] Step 2: Extract the representation vector of each attribute representation matrix, and based on the original distribution of the ad slots on the target display screen, obtain the multidimensional features of each ad slot, wherein the multidimensional features include: main representation features and auxiliary representation features of the other ad slots for the corresponding ad slots;

[0067] Step 3: Match the historical change baseline that matches each multi-dimensional feature from the historical delivery database, and rely on the historical change baseline and combine it with the added virtual ads to predict the delivery advantage of the corresponding ad slot within the preset time period;

[0068] Step 4: Perform time alignment based on the advantages of all ad placements, determine the ad placement prominence coefficient for each aligned time point, and lock the time points where the prominence coefficient is greater than the preset coefficient as the cut-off points to optimize the ad placement strategy.

[0069] In this embodiment, the database is a structured data storage system used to store historical operating data, interaction data, and related attribute information of all advertising spaces on the target display screen. For example, the advertising log library of the display screen on the 3rd floor of the shopping mall contains the playback records, user interaction data, virtual advertising simulation data, etc. of each advertising space.

[0070] In this embodiment, the target display screen is a physical display device used to display advertising content. It is the carrier of the advertising space and can be located indoors, such as a shopping mall screen, or outdoors, such as an outdoor large screen.

[0071] In this embodiment, an ad slot is an independent advertising display area divided on the target display screen. Each area can display advertising content independently and has independent attribute characteristics and interactive data. For example, in a two-screen display, the left side of the first screen is ad slot A and the right side is ad slot B. The physical boundaries of each ad slot are determined by screen partitioning technology and associated with an independent playback control module.

[0072] In this embodiment, the multidimensional attribute information reflects a multidimensional data set of ad space characteristics, covering multiple aspects such as playback behavior, traffic quality, and virtual ad performance. For example, the multidimensional attribute information of ad space A includes: 20 active playback times / day, 15 passive trigger times / day, an average of 5 people per passive playback, 3 virtual ads, and 20 virtual ad clicks.

[0073] In this embodiment, the pre-ad proactive playback information is the relevant data of the ad slot automatically rotating ads when no user triggers it, reflecting the proactive exposure of the ads. For example, ad slot A automatically rotates a milk tea ad from 9:00 to 22:00 every day, playing a total of 30 times for a total duration of 60 minutes, of which 10 times are played during the prime time period from 18:00 to 20:00.

[0074] In this embodiment, the pre-ad passive playback information is the relevant data of the ad space playing the ad when the user triggers such as approaching the screen or touching the screen. It reflects the user's active attention behavior. For example, when the user approaches ad space B, the screen senses and triggers the playback of an ad for a certain sports shoe. This is triggered 25 times in total, with an average playback duration of 45 seconds. The user touches the screen 8 times. Specifically, the user's trigger behavior is detected by devices such as infrared sensors and touch screens, and data such as trigger time, playback duration, and interaction actions are recorded.

[0075] In this embodiment, the traffic volume is the real-time number of viewers around the ad slot during each passive playback, reflecting the reach and quality of passive exposure. For example, when ad slot B is passively played for the first time, there are 3 people watching it; the second time, there are 7 people watching it, with an average traffic of 5 people / play.

[0076] In this embodiment, the playback status of virtual advertisements refers to the playback data of virtual advertisements during the delivery process, including the number of playbacks, simulated interaction volume, audience dwell time, etc., which are used to assist in the prediction of real advertising effects. For example, the virtual advertisement in ad slot C was played 15 times, simulated clicks 9 times, and the average simulated dwell time was 8 seconds. Among them, the simulated interaction of young female users accounted for 60%.

[0077] In this embodiment, quantization is the process of converting non-numerical or multidimensional attribute information with different dimensions into a unified numerical form to eliminate the impact of data differences on subsequent analysis. For example, the number of active playbacks is 30, the maximum value is 37, and the standardized value is 0.8. The number of passive triggers is 25, the maximum value is 28, and the standardized value is 0.9. Algorithms such as Z-score standardization are used to output quantized values ​​in the range of 0 to 1.

[0078] In this embodiment, the attribute representation matrix is ​​a set of quantified attributes of all ad slots presented in matrix form.

[0079] In this embodiment, the original distribution is the physical arrangement of the advertising spaces on the target display screen.

[0080] In this embodiment, the multidimensional features are a set of features that comprehensively reflect the characteristics of the ad placement itself and the influence of other ad placements. They include primary representation features and secondary representation features. For example, the multidimensional features of ad placement A include: primary representation features: active playback intensity and passive interaction depth; secondary representation features: visual interference degree of ad placement B and traffic diversion rate of ad placement D.

[0081] In this embodiment, the historical advertising database is a database that stores all advertising placement data and performance metrics over a past period of time. It is used to extract historical patterns and baselines. For example, it stores the advertising placement records of shopping mall display screens over the past 6 months, including daily playback data, conversion results, user feedback, etc.

[0082] In this embodiment, the historical change baseline is the effect change pattern of advertising cases with similar multidimensional features to the current ad slot in the historical advertising database, which serves as a reference benchmark for prediction. For example, the current multidimensional features of ad slot A are: high passive interaction and medium traffic. It matches 3 advertising cases with similar features in the past, and its effect baseline is: the advertising advantage shows an upward trend from 18:00 to 20:00, with a peak at 19:30. Specifically, the K nearest neighbor algorithm is used to match similar historical features, and the effect change curve over time is extracted through a sliding window to form the historical baseline.

[0083] In this embodiment, the preset time period is a pre-set time range used to predict the advantages of ad placement. It can be adjusted according to the needs of the scenario, such as 1 day or 1 week.

[0084] The placement advantage is the potential effect of an ad placement within a preset time period. It quantifies the comprehensive value of exposure quality, conversion probability, and other factors. Time alignment processing unifies the placement advantage prediction results of different ad placements onto the same time axis. The aligned time point is a unified time node obtained after time alignment processing. Each node corresponds to the start or end time of a time unit.

[0085] In this embodiment, the ad placement prominence coefficient is a quantitative indicator that measures the prominence of ads placed at aligned time points, taking into account factors such as placement advantages, interference factors, and user activity.

[0086] In this embodiment, a preset coefficient is used to determine whether the prominence coefficient meets the threshold. It is set by historical performance data or business needs, and time points that are higher than this value are considered high-quality delivery nodes.

[0087] In this embodiment, the cutting point is the alignment time point where the highlighting coefficient is greater than a preset coefficient.

[0088] In this embodiment, the optimization of the advertising placement strategy is a process of improving the placement effect by adjusting the time allocation, content selection, frequency, etc. of advertising placement based on the cut point. For example, for the cut point of ad position B, the optimization strategy is to play the catering ad 3 times from 12:30 to 13:00 and the clothing ad 2 times from 18:30 to 19:00, and match it with highly attractive creative materials.

[0089] The beneficial effects of the above technical solution are: by comprehensively quantifying the characteristics of ad placements with multi-dimensional attributes, mining the correlation between ad placements, integrating historical data with the advantages of virtual advertising prediction, and accurately optimizing based on time nodes, the targeting and effectiveness of advertising can be effectively improved, resource waste can be reduced, and the anti-interference ability and adaptability of the strategy can be enhanced through virtual advertising and interference processing.

[0090] This invention provides a method for optimizing an advertising delivery strategy, which involves acquiring additional virtual advertisements, including:

[0091] Acquire the generated creatives, visual effects, and placement environment for each ad placement based on the target display screen, and construct the placement bias for each ad placement;

[0092] Determine the first sequence of historical order-brushing attributes and the second sequence of targeting bias for each ad placement;

[0093] Extract the feature parameters of the first sequence and the second sequence, determine the connection relationship between the feature parameters, and add the initial advertisement to the corresponding advertisement position;

[0094] The initial advertisement is split into single frames, a directed graph of order-brushing elimination is constructed for each split frame, and the elimination probability is obtained;

[0095] The initial advertisement is divided using an interactive window, and the elimination characteristics of each sub-advertisement are determined by combining the elimination probability.

[0096] Linear detection is performed on all eliminated features to extract the best features. The best features are combined with the global coverage ratio and trapping strength to optimize the initial advertisement and obtain the virtual advertisement.

[0097] In this embodiment, the generated materials for the ad space are used to create the original materials for the ad content, including basic content elements such as images, video clips, and text copy. Visual effects refer to the visual characteristics of the ad content, including perceptible visual attributes such as color saturation, the proportion of dynamic elements, font size, and image complexity. The placement environment refers to the physical scene and surrounding conditions of the ad space, including geographical location (e.g., a women's clothing section in a shopping mall, a subway transfer passage), pedestrian flow characteristics (e.g., predominantly young women, high pedestrian density during morning and evening rush hours), and lighting conditions (e.g., ample natural light, nighttime artificial lighting).

[0098] The placement bias is determined based on the generated materials, visual effects, and placement environment of the ad placement, indicating the type of advertisement or content direction that the ad placement is more suitable for: such as product category and style. For example, the placement bias is: women's clothing and beauty advertisements, with a visual style that is fashionable and youthful, and content that focuses on new product promotion.

[0099] Historical order-brushing attributes are the characteristics of past fraudulent operations on ad placements, such as machine clicks and malicious repeated views, including the frequency of order-brushing, the time period of order-brushing, and the device characteristics of fake interactions.

[0100] The first sequence is a sequence of historical order-brushing attribute data for ad slots arranged in chronological order. It is used to reflect the time-varying pattern of order-brushing behavior. For example, the first sequence of ad slot A is: 5 times in January 2024, 3 times in February, 0 times in March, and 6 times in April (monthly as time unit). That is, the quantitative values ​​of historical order-brushing attributes are sorted by time and stored as sequence data using timestamps as indexes.

[0101] The second sequence is a time-series data sequence of ad placement bias, used to reflect the time-varying patterns of the bias. For example, the second sequence of ad placement A is: January 2024: bias towards women's clothing; February: women's clothing and beauty; March: beauty; April: women's clothing. In other words, the bias tags are sorted by time to form a sequence with the same time granularity as the first sequence.

[0102] Feature parameters are key quantitative indicators extracted from the first and second sequences to describe the core features of the sequence. For example, the feature parameters extracted from the first sequence are: an average of 4.2 times of fraudulent orders per month and the peak time for fraudulent orders is 3 a.m.; those extracted from the second sequence are: the frequency of the women's clothing tag is 75% and the bias switching cycle is 2 months. Specifically, the mean, peak value, frequency and other parameters are extracted through time series analysis tools to quantify the sequence features.

[0103] The connection relationship is the correlation between the feature parameters of the first sequence and the second sequence, reflecting the mutual influence between order-faking behavior and the targeting bias. For example, the analysis found that when the targeting bias is women's clothing, the number of order-faking times is 30% higher than that of beauty products, which is a positive correlation connection relationship. Specifically, the correlation strength between feature parameters is calculated through correlation analysis and causal inference model, and the connection relationship matrix is ​​output.

[0104] The initial advertisement is based on the connection between the first and second sequences. The initially generated virtual advertisement has not been optimized for fraudulent orders. Its content is matched with the ad placement bias and is used for subsequent optimization. For example, based on the connection relationship that the fraudulent order rate is high when women's clothing is biased, an initial advertisement is generated for ad placement A, such as a virtual women's clothing advertisement, which includes basic anti-fraudulent order indicators.

[0105] Single-frame splitting involves dividing the initial video advertisement into static image frames (single frames) at time intervals, making it easier to analyze the characteristics of fraudulent orders frame by frame.

[0106] The directed graph for eliminating fraudulent orders uses individual frames as nodes, with directed edges between nodes representing fraudulent order characteristics, such as fake click locations and abnormal dwell times, and their transmission relationships between frames. This is used to identify the propagation patterns of fraudulent order behavior. For example, if the nodes are frames 1 to 5, and the edges are frames 2 to 3, it means that the fake click location that appears in frame 2 is repeated in frame 3. The directed graph can locate the frame sequence in the fraudulent order set. Specifically, a directed graph is constructed using graph theory tools, where the node attributes are the fraudulent order characteristics of the frame, and the edge weights are the feature transmission probabilities.

[0107] Elimination probability is the probability that a single frame can effectively eliminate the interference of fraudulent clicks during the delivery. For example, elimination probability EP = (1 - percentage of abnormal clicks) × (1 - device repetition rate) × click location dispersion.

[0108] Interactive windows are segments of time in the initial advertisement where users may interact, such as clicking or staying on the screen. The duration of the interaction is usually determined based on historical interaction data.

[0109] Sub-ads are ad segments corresponding to each interactive window, and are analyzed for elimination features as independent units.

[0110] Elimination features are the characteristics that can be used to identify and exclude fraudulent behavior in sub-ads, such as the location distribution of abnormal clicks, the dispersion of interaction duration, and the repetition rate of device fingerprints. For example, the elimination features of sub-ad 2 are: the click locations are concentrated in non-ad content areas, and the frequency of 8 clicks within 10 seconds on the same device is abnormal.

[0111] Linearity detection eliminates linear correlations between features by analyzing whether the trends of feature A and feature B are consistent, thus selecting independent and effective features. The linear correlation between features is calculated using the Pearson correlation coefficient, and redundant features with a correlation coefficient > 0.7 are eliminated.

[0112] The best feature is the feature that contributes the most to identifying fraudulent order behavior after linear detection. For example, the feature that contributes the most is related to the highest accuracy and the lowest false positive rate.

[0113] Global coverage ratio is the proportion of historical fraudulent behavior that the best feature can cover for the ad placement. It reflects the overall recognition range of fraudulent behavior by the feature. For example, if the IP overlap rate of the best feature is >60%, it can cover 85% of the historical fraudulent behavior of the ad placement. The trapping intensity is the sensitivity of the best feature to fraudulent behavior, that is, the proportion of fraudulent behavior that can be accurately captured. The higher the intensity, the stronger the trapping ability.

[0114] Virtual ads are simulated ads obtained by combining the best features of the initial ads with global coverage ratio and trapping strength. They are used to assist in evaluating the effect of real ad placements and have stronger resistance to fraudulent orders. For example, the optimized virtual women's clothing ad retains the original visual style, but embeds IP overlap rate detection features with a coverage ratio of 85% and a trapping strength of 0.9, which can effectively eliminate fake interaction data.

[0115] The beneficial effects of the above technical solution are: by combining the characteristics of ad placement with historical patterns of fraudulent orders, virtual ads with strong anti-fraud capabilities are generated. This not only matches the placement bias of ad placements but also accurately identifies fake interactions, providing a reliable reference for evaluating the effectiveness of real ad placements and improving the accuracy of ad strategy optimization.

[0116] This invention provides a method for optimizing advertising delivery strategies, which involves obtaining multi-dimensional attribute information of each ad slot on a target display screen from a record database and performing quantification to construct an attribute representation matrix, including:

[0117] The multidimensional attribute information is input into a quantization processing model that matches the target display screen to obtain the multidimensional output vector of the corresponding ad position. At the same time, the user behavior dynamic features of each ad position and the spatial correlation parameters of adjacent ad positions are obtained as extended dimensions of the multidimensional attribute information and quantized to obtain the extended vector.

[0118] Based on the mapping relationship between each extended dimension and each dimension type in the multidimensional attribute information, the mapping priority is determined, and a mapping path is generated;

[0119] Based on the mapping path and the set of ad variations for each dimension type, a topology network is constructed.

[0120] The topology network is analyzed to determine the parameters to be retained in the extended dimension and the influence factors of each retained parameter on the corresponding original elements under the mapping relationship. The new vector is then obtained.

[0121] Map all new vectors to a high-dimensional feature space to generate attribute representation matrices.

[0122] In this embodiment, the quantization processing model is a machine learning model customized for the hardware characteristics of a specific target display screen, such as size, resolution, and deployment scenario, such as shopping malls, subways, and user characteristics. It is used to convert unstructured or multidimensional attribute information of different dimensions into standardized numerical vectors. It is obtained by training the model based on the historical data of the target display screen. For example, the quantization processing model for the LED display screen on the 3rd floor of the shopping mall is based on the user being mainly young women and the screen size being 55 inches. It uses a random forest regressor and outputs standardized multidimensional vectors after inputting the original attribute data.

[0123] A multidimensional output vector is a standardized numerical vector obtained after multidimensional attribute information is quantized and transformed by a model. For example, after the multidimensional attribute information of ad slot A is processed by the model, the multidimensional output vector {0.82, 0.65, 0.58, 0.42, 0.36} is obtained, which correspond to the quantified values ​​of active playback count, passive trigger count, traffic volume, number of virtual ads, and number of virtual ad clicks, respectively.

[0124] User behavior dynamic features are real-time dynamic behavioral data during the interaction between users and ad placements, including the rate of change in user gaze duration, fluctuations in click frequency, and the direction of movement of the dwell trajectory, which are behavioral characteristics that change over time.

[0125] The spatial correlation parameters of adjacent ad positions reflect the physical spatial correlation characteristics between the current ad position and adjacent ad positions, including spatial distance, such as 2 meters, visual interference, such as the color contrast of adjacent ad positions, 0.7, and user attention shift probability, such as the probability of looking from ad position A to ad position B, 30%.

[0126] The extended dimensions are new dimensions relative to the original multidimensional attribute information, namely the dynamic characteristics of user behavior and the spatial correlation parameters of adjacent ad positions, which are used to supplement the deeper characteristics of ad positions, such as real-time user feedback and spatial interaction effects.

[0127] The extended vector is a standardized vector obtained after quantifying the extended dimension. It has the same dimensions as the multidimensional output vector, both ranging from 0 to 1, which facilitates subsequent fusion analysis. For example, the extended vector of ad slot A is {0.62, 0.45, 0.38, 0.52}, which correspond to the quantified values ​​of gaze duration change rate, click frequency fluctuation value, distance from adjacent ad slots, and visual interference, respectively.

[0128] The correlation between the extended dimensions and the original multidimensional attribute dimensions includes positive correlation, negative correlation, or no correlation. For example, the mapping relationship between the rate of change of gaze duration and the number of passive playbacks is positively correlated with a correlation coefficient of 0.7; the mapping relationship between visual interference and active playback effect is negatively correlated with a correlation coefficient of -0.6.

[0129] Mapping priority is determined by the strength of the mapping relationship, i.e., the absolute value of the correlation coefficient or the degree of influence on the ad placement effect, and determines the priority order of the extended dimension and the original dimension. For example, the mapping priority order is: gaze duration change rate → passive playback count with priority of 0.9 > visual interference degree → active playback effect with priority of 0.7 > click frequency fluctuation value → traffic size with priority of 0.5.

[0130] The mapping path is a sequence of associations between the extended and original dimensions arranged according to mapping priority. It reflects the influence transmission path from the extended dimension to the original dimension, such as the rate of change in gaze duration → number of passive playbacks → traffic volume. For example, the mapping path of ad slot A is: rate of change in gaze duration → number of passive playbacks → visual interference → number of active playbacks → click frequency fluctuation value → virtual ad clicks.

[0131] Each dimension type's ad change set is a data collection of changes over time for each original or extended dimension, reflecting the dynamic trends of dimension characteristics, such as daily changes in passive play counts and hourly fluctuations in the rate of change in gaze duration. For example, the ad change set for passive play counts is: Monday 15 times → Tuesday 18 times → Wednesday 22 times → Thursday 17 times; the change set for gaze duration rate is: 10:00 0.3 → 12:00 0.6 → 14:00 0.5.

[0132] A topological network is a network model constructed using dimension types as nodes, mapping paths as edges, and ad change sets as node attributes. It is used to intuitively display the relationships and dynamic changes between various dimensions.

[0133] Topology analysis is a quantitative analysis of the node association strength, path transmission efficiency, and node attribute change trends of a topology network. It is used to screen extended dimension parameters that have a significant impact on the original dimensions. For example, the analysis found that the correlation strength between the gaze duration change rate and the number of passive playbacks is significantly higher than other correlations, and its change set is highly consistent with the change trend of the number of passive playbacks.

[0134] The parameters retained in the extended dimensions are the specific parameters of the extended dimensions that have a significant impact on the original dimensions, selected after topological network analysis. For example, in the dynamic features of user behavior in the extended dimensions, only parameters with a gaze duration change rate > 0.5 are retained, while parameters with a click frequency fluctuation value < 0.3 are removed.

[0135] The impact factor quantifies the degree of influence of retained parameters on the specific values ​​of original dimensional elements such as the number of passive plays, and its value ranges from -1 to 1.

[0136] Element adjustment involves adjusting the values ​​of corresponding elements in the original multidimensional output vector based on the impact factor, so that the adjusted elements more accurately reflect the influence of the extended dimensions. For example, if the element value of the passive playback count in the original multidimensional output vector is 0.6, after adjustment by the retained parameter gaze duration change rate (impact factor 0.7), the new value = 0.6 × (1 + 0.7) = 1.02 (1.0 after standardization). The formula used is: Adjusted element value = Original value × (1 + Impact factor).

[0137] The new vector is a vector obtained after element adjustment, incorporating the effects of the expanded dimensions. For example, the original multidimensional output vector of ad slot A is {0.82, 0.6, 0.58, 0.42, 0.36}. After element adjustment, the new vector is {0.82, 1.0, 0.65, 0.42, 0.45}, where the passive play count and traffic volume elements are increased due to the effects of the expanded dimensions.

[0138] A high-dimensional feature space is a feature space with a dimension higher than that of the original vector, such as expanding from 5 dimensions to 10 dimensions. It is used to capture more complex feature interaction relationships in the new vector, such as the cross feature of passive playback count × gaze duration change rate. For example, mapping a 5-dimensional new vector to a 10-dimensional high-dimensional space adds features such as: passive playback count × gaze duration change rate, active playback count × visual interference degree, etc.

[0139] The attribute representation matrix is ​​a matrix formed by arranging the high-dimensional mapping results of multiple ad slots in rows. Rows represent ad slots, columns represent high-dimensional features, and elements are the quantified values ​​of the corresponding features. It is the core input for subsequent feature extraction and strategy optimization. Taking the attribute representation matrix of 3 ad slots as an example, it simplifies to 10 columns:

[0140]

[0141] The beneficial effects of the above technical solution are: by introducing dynamic features of user behavior and spatial correlation parameters as extended dimensions, and by combining mapping relationships and topological network optimization of the original attribute vector, the resulting attribute representation matrix can more comprehensively and accurately capture the core features of the ad placement, providing a more reliable quantitative basis for the optimization of subsequent ad placement strategies, and improving the targeting and effectiveness of the strategies.

[0142] This invention provides a method for optimizing an advertising placement strategy, which relies on the original distribution of ad slots on the target display screen to obtain multi-dimensional features of each ad slot, including:

[0143] The attribute representation matrix is ​​subjected to nonlinear dimensionality reduction to extract a deep principal representation vector containing high-order feature interactions;

[0144] A spatial association graph of ad placements is constructed based on the physical coordinate distribution of the target display screen, where nodes are ad placements and edge weights are the visual attention shift probability and spatiotemporal co-occurrence frequency of adjacent ad placements.

[0145] Graph neural networks are used to propagate features in a spatial relational graph. Noise features of nodes and path segments between nodes are extracted during the propagation process to construct a noise sequence and extract local and global significant noise factors.

[0146] The feature propagation process is noise-eliminating based on all local and global significant noise factors, generating spatial correlation feature vectors for each ad slot as auxiliary representation features. In the feature propagation process, a distance attenuation coefficient is introduced so that the feature influence of adjacent ad slots decreases linearly with the increase of physical distance.

[0147] By combining the characteristic fluctuation sequence of the ad placement within a preset historical period, a time-series dynamic feature vector is extracted as a supplementary auxiliary representation feature.

[0148] Based on deep master representation vectors, spatial correlation feature vectors, and temporal dynamic feature vectors, multidimensional features of corresponding ad slots are generated.

[0149] In this embodiment, nonlinear dimensionality reduction is the process of mapping a high-dimensional attribute representation matrix to a low-dimensional space through nonlinear transformation, preserving the nonlinear structure and key features of the data. For example, in ad slot B, the higher-order interaction between the passive trigger quantization value of 0.9 and the traffic size quantization value of 0.8 is manifested as the virtual ad click-through rate increasing by 2 times when their product is greater than 0.7. This correlation cannot be captured by linear relationships.

[0150] In this embodiment, the deep principal representation vector is a low-dimensional vector that reflects the core attributes and high-order feature interactions of the ad slot after nonlinear dimensionality reduction. For example, the deep principal representation vector {0.85, 0.72, 0.68} is obtained after dimensionality reduction of the 5-dimensional attribute representation matrix of ad slot B.

[0151] In this embodiment, the physical coordinate distribution of the target display screen refers to the specific location coordinates of the advertising space on the target display screen, reflecting the spatial arrangement of the advertising spaces.

[0152] In this embodiment, the ad space association graph is a graph structure constructed with ad spaces as nodes and spatial association relationships as edges. It is used to visualize the spatial interaction between ad spaces. For example, in the spatial association graph of 3 ad spaces, the nodes are A, B, and C. There is an edge between A and B, which means they are adjacent. There is an edge between A and C, which means they are vertically opposite. There is no edge between B and C, which means they are not adjacent.

[0153] In this embodiment, a node is the basic unit representing a single ad space in the ad space spatial association graph. Each node corresponds to the identifier and core attributes of an ad space. The edge weight is the weight value of the edge connecting two nodes in the spatial association graph, which comprehensively reflects the spatial association strength between the two ad spaces. Edge weight = 0.6 × visual attention shift probability + 0.4 × spatiotemporal co-occurrence frequency.

[0154] In this embodiment, the visual attention shift probability is the probability that a user will shift their gaze from one advertisement to another, reflecting the degree of visual attraction between the two advertisements.

[0155] In this embodiment, the spatiotemporal co-occurrence frequency is the frequency at which two ad slots are simultaneously viewed by the same group of users within a unit of time. For example, if the same user views both ad slots A and B simultaneously 20 times within one hour, it reflects the degree of audience overlap between the two ad slots.

[0156] In this embodiment, the graph neural network is a neural network specifically designed to process graph structure data. It can learn the mutual influence between nodes through the information transmission between nodes and edges. Specifically, it constructs a GNN model, inputs the node features and edge weights of the spatial association graph, and sets three hidden layers for feature learning.

[0157] In this embodiment, feature propagation is the process by which node features in a GNN are passed to neighboring nodes through edges, enabling each node to integrate the feature information of its neighbors. For example, A's features are passed to B, and B's features are also passed to A.

[0158] A propagation path is the path through which a feature is transmitted from one node to another, such as A→B→C, meaning that the feature of A first reaches B, and then from B to C. Node noise features are the interference information carried by a single node itself during feature propagation, such as outliers caused by data collection errors or accidental false interactions. For example, a passive trigger data of ad slot A might be overestimated due to sensor malfunction, misrecording 10 interactions as 20. This outlier forms node noise features during propagation. Path segment noise features are the interference information introduced by a feature in a certain segment of the propagation path. For example, the visual attention shift probability between A and B is actually 0.6, but due to statistical error, it is miscalculated as 0.8, causing the feature received by B from A to be over-amplified, forming path segment noise.

[0159] A noise sequence is a sequence of node noise features and path segment noise features arranged in the order of the propagation path, reflecting the distribution pattern of noise during propagation. For example, the noise sequence of the path A→B→C is: node A noise → A→B path segment noise → node B noise → B→C path segment noise → node C noise. In this case, the noise features are spliced ​​together in the time order of the propagation path to form a one-dimensional sequence, such as: {0.1,0.05,0.12,0.08,0.09}.

[0160] In this embodiment, the local significant noise factor is a quantified value of noise intensity that only affects a single propagation path or local node. For example, if the noise on a certain path only interferes with A and B, but does not affect C, the variance of the noise on the local path is calculated. Generally, the larger the variance, the more significant the local impact. After normalization, the factor is between 0 and 1, specifically:

[0161] ;

[0162] Where p is the identifier of a specific propagation path or local node cluster; Let p be the total number of nodes; Let p be the set of noise eigenvalues ​​for all nodes. Let p be the set of noise feature values ​​for all path segments; Based on The variance; Based on The peak value; Lp is the weighting coefficient; Lmax is the physical length of p; Lmax is the maximum physical length of all paths on the target display screen. For distance attenuation, LSp is the local significant noise factor, with a value between 0 and 1.

[0163] The global significant noise factor is a quantified value of the noise intensity affecting multiple paths or nodes in the entire spatial association graph (such as a systematic error causing all ad placements to have excessively high passive trigger data). For example, if a widespread sensor malfunction occurs during a certain period, causing the passive trigger data of A, B, and C to be 20% higher, this noise affects the entire graph, and the global significant noise factor is 0.5.

[0164] ;

[0165] Where Nall is the set of noise from all nodes and path segments in the overall topology network; Avg(Nall) is the average value of global noise; Dnet is the network density; Entropy(Nall) is the information entropy of global noise; Ccow is the noise coverage; and GSF is the global significant noise factor, with a value between 0 and 1.

[0166] Noise cancellation processing is based on local and global significant noise factors. It filters or corrects noise in the feature propagation process to reduce interference. For example, for noise with a local noise factor of 0.3 in the A→B path, the feature value is corrected by using the formula: feature value = original value × (1 - noise factor). For noise with a global noise factor of 0.5, the deviation value is uniformly subtracted, such as the 20% artificially high part.

[0167] The spatial association feature vector is a vector that integrates the features of neighboring ad spaces after noise removal, reflecting the spatial association between the ad space and its surrounding environment.

[0168] Auxiliary representation features are spatial correlation feature vectors used as supplementary features. The auxiliary deep principal representation vectors provide a more comprehensive description of the ad placement. The deep principal representation reflects its own characteristics, while the auxiliary features reflect the surrounding influences. For example, the deep principal representation vector {0.8, 0.6, 0.5} (its own attributes) of ad placement A, plus the spatial correlation feature vector {0.65, 0.72, 0.58} (the surrounding influences), together constitute a more complete feature description.

[0169] The distance decay coefficient is a coefficient that controls how the influence of an ad placement's features weakens as physical distance increases. The farther the distance, the smaller the coefficient, and the weaker the feature influence. The distance decay coefficient is calculated as: 1 - current distance / maximum distance.

[0170] The preset historical period is a pre-defined historical time range used to analyze changes in ad placement characteristics.

[0171] The characteristic fluctuation sequence is the sequence of changes in the characteristics of an ad placement over a preset historical period, such as the daily passive trigger quantification value or the weekly changes in virtual ad clicks.

[0172] The time-series dynamic feature vector is a vector extracted from the feature fluctuation sequence that reflects the trend of ad placement features over time, such as an upward trend or periodic fluctuations. For example, the time-series dynamic feature vector of ad placement B is {0.8, 0.3, 0.6}, which corresponds to high fluctuations on weekdays, stability on weekends, and an average growth trend over 7 days, respectively. The trend term, periodic term, and residual term are extracted using LSTM or time series decomposition STL, and then quantized to form a vector.

[0173] Supplementary auxiliary representation features are additional auxiliary features that supplement the temporal dynamic feature vector to describe the temporal dimension changes of the ad placement. Among them, the deep main representation reflects static attributes, spatial correlation reflects spatial influence, and supplementary auxiliary features reflect temporal dynamics.

[0174] Multidimensional features are comprehensive feature vectors that integrate deep master representation vectors, spatial correlation feature vectors, and temporal dynamic feature vectors, fully reflecting the multidimensional attributes of ad placements.

[0175] The beneficial effects of the above technical solution are: by extracting the core features of the ad placement through nonlinear dimensionality reduction, combining graph neural networks to mine spatial correlations and eliminate noise, and incorporating temporal dynamic changes, the generated multidimensional features can comprehensively capture the ad placement's own attributes, surrounding influences and time patterns, providing a more accurate quantitative basis for subsequent placement advantage prediction and strategy optimization, and effectively improving the targeting and effectiveness of advertising.

[0176] This invention provides a method for optimizing an advertising placement strategy, relying on the historical change baseline and combining it with added virtual advertising to predict the placement advantage of corresponding ad slots within a preset time period, including:

[0177] The historical change baseline is decomposed into multiple sub-baseline sequences according to the time granularity. Each sub-baseline sequence corresponds to a time unit within a preset time period. The trend characteristic parameters and fluctuation thresholds of each sub-baseline sequence are extracted.

[0178] The playback data of the newly added virtual advertisements is analyzed in layers to obtain the interaction conversion rate and content retention rate of virtual advertisements in different user profile groups, and a two-dimensional profile matrix of virtual advertisements is constructed.

[0179] The elements in each dimension of the two-dimensional profile matrix are sorted in descending order according to their values, and a sequence number group is constructed for each remaining user except for the last user in the sort.

[0180] A difference matrix is ​​constructed based on the difference between two adjacent rows of the matrix after descending order. The results of each column in the difference matrix are clustered and classified, and the mode range of each column is determined. The minimum value in the mode range is used to replace the element values ​​before the mode range, and the maximum value in the mode range is used to replace the element values ​​after the mode range to obtain a new two-dimensional matrix. The two-dimensional arrays of the remaining users are then obtained.

[0181] The effect mapping vector is obtained by relying on the two-dimensional array, the index group and the original array of each remaining user;

[0182] The fusion weights of the historical change baseline and the effect mapping vector are dynamically adjusted based on the time decay function. The weight of the effect mapping vector increases exponentially with the increase of real-time interactive data, while the weight of the historical change baseline decreases linearly with the extension of the time interval.

[0183] The fused feature vector and environmental variable parameters for a preset time period are input into the delivery advantage prediction model, and the delivery advantage quantification value for each time unit is output.

[0184] The quantitative value of the delivery advantage in continuous time units is smoothed to generate a delivery advantage curve that includes local peaks and global trends. The area under the curve and the duration of the peak are used as the final evaluation indicators of the delivery advantage.

[0185] In this embodiment, the time granularity is the time unit for splitting historical change baselines. It is used to decompose continuous baselines into discrete subsequences. The granularity is selected according to the length of a preset time period, such as selecting 1 hour for 1 day or 1 day for 1 week, to ensure that the number of subsequences is appropriate.

[0186] Sub-baseline sequences are local sequences obtained by splitting historical change baselines by time granularity, with each sequence corresponding to baseline data for one time unit.

[0187] The preset time period is a pre-defined future time range used to predict the advantages of deployment, such as 2 days, which corresponds to the time unit of the sub-baseline sequence, and the time unit supports dynamic adjustment.

[0188] Trend characteristic parameters are quantitative indicators that describe the changing trend of the sub-baseline sequence, including the upward / downward slope, average growth rate, number of inflection points, etc. The slope is calculated by linear regression, the growth rate is calculated by the difference method, and the point where the derivative is 0 is identified as an inflection point. After quantification, a parameter set is formed.

[0189] The fluctuation threshold is the maximum allowable range of data fluctuation in the sub-baseline sequence. It is used to determine whether the sequence is stable. Fluctuations exceeding the threshold are considered abnormal. Specifically, the standard deviation of the sub-baseline sequence is calculated, and three times the standard deviation is used as the fluctuation threshold, which conforms to the 99.7% confidence interval of the normal distribution.

[0190] Layered analysis divides virtual ad playback data into different layers based on user characteristics such as age, gender, and consumption habits, and analyzes the effectiveness of each layer separately. For example, virtual ad playback data can be divided into four layers based on age and gender, and the interaction conversion rate can be calculated separately for each layer.

[0191] A user profile is a group of users who share common characteristics, including age, gender, behavioral habits, interests, and preferences.

[0192] Interaction conversion rate is the percentage of virtual ad users who interact with the ad, such as those who click or stay for more than 3 seconds, out of the total number of viewers in that group, reflecting the group's interest in the ad.

[0193] Content retention rate is the percentage of users who can recall the core content of a virtual advertisement after watching it, reflecting the effectiveness of the advertisement's dissemination.

[0194] A two-dimensional user profile matrix is ​​a matrix constructed with user profile groups as rows and interaction conversion rate and content retention rate as columns. The elements are the quantitative values ​​of the corresponding groups, i.e., 0 to 1. For example, a two-dimensional user profile matrix with 4 groups:

[0195]

[0196] The dimension of the two-dimensional profile matrix is ​​the column dimension of the matrix. Here, it refers to the two indicator dimensions of interaction conversion rate and content retention rate. For example, the dimensions of the matrix are interaction conversion rate (column 1) and content retention rate (column 2). The descending order sorts the matrix rows according to the element value of each dimension from largest to smallest, which is convenient for analyzing the differences in group effects.

[0197] The sequence group is the remaining rows after excluding the last row in descending order, and is assigned the sequence number 1, 2, 3, etc. according to the sorting order. It is used to mark the relative priority of the group.

[0198] The difference matrix is ​​calculated by subtracting the values ​​of corresponding elements in adjacent rows from the values ​​of the previous row to the values ​​of the next row in a descending matrix. It reflects the performance differences between groups. For example, if the interaction conversion rate of the remaining three rows after descending order is {0.3, 0.25, 0.2} and the content retention rate is {0.7, 0.5, 0.6}, the difference matrix would be:

[0199]

[0200] Clustering classification involves grouping the difference results in each column of the difference matrix to identify patterns of difference, such as significant differences and minor differences. The K-means clustering algorithm is used to group the difference results in each column, and the optimal K value is determined by the silhouette coefficient.

[0201] The mode range is the interval containing the most frequent value in each column of differences. For example, if the mode is 0.05, the range is 0.04 to 0.06, reflecting the typical difference in that dimension.

[0202] Equivalent replacement replaces elements before the range with the minimum value of the range and elements after the range with the maximum value, simplifying the data while preserving the core differences.

[0203] The new two-dimensional matrix is ​​obtained by equivalent replacement. The elements are simplified interaction conversion rate and content retention rate values, which are easier to calculate later. For example, after the replacement, the interaction conversion rate of women aged 20 to 30 in the original matrix is ​​adjusted from 0.3 to 0.32, forming the new matrix.

[0204] The two-dimensional array is a row vector corresponding to each remaining user group in the new two-dimensional matrix. It contains two elements: interaction conversion rate and content retention rate. The elements of the new two-dimensional matrix are extracted row by row and stored as nested lists, such as {0.32,0.71} and {0.27,0.51}.

[0205] The original array is the original row vector corresponding to the remaining user groups in the two-dimensional portrait matrix before it is sorted in descending order, that is, the original values ​​before sorting and replacement.

[0206] The effect mapping vector is generated by fusing the two-dimensional array of the remaining user groups, the index group, and the original array to comprehensively reflect the effect of virtual advertising on different groups. Specifically:

[0207] Interaction conversion dimension:

[0208] ;

[0209] in, J1 represents the interaction conversion dimension component; J1 is the adjusted interaction conversion rate. R is the original interaction conversion rate; S1 is the user group number; wr is the confidence score of the group interaction data, ranging from 0 to 1; β is the number weight baseline value; β is the number decay coefficient; θ1 is the difference penalty coefficient.

[0210] Content memory dimension:

[0211] ;

[0212] in, A1 represents the content memory dimension component; A2 represents the adjusted content memory retention rate. denoted as the original content retention rate; T is the time interval after content exposure; M is the memory enhancement coefficient; and k1 is the memory decay coefficient.

[0213] in, , It is mapped to the [0, 1] interval and is directly used for subsequent fusion calculations with historical baselines.

[0214] The time decay function is used to dynamically adjust the weights of historical and real-time data, so that the weight of old data decreases over time and the weight of new data increases. For example, the time decay function of the historical change baseline is: weight = 1 - t / T, where t is the time interval and T is the total period, such as t = 5 days or T = 30 days.

[0215] The fusion weight is the weight ratio of the historical change baseline and the effect mapping vector during fusion, with a total of 1, used to balance the influence of the two on the prediction results.

[0216] Real-time interactive data volume is the total amount of cumulative interactive data of virtual advertisements at the current moment, such as the number of clicks and the total dwell time, reflecting the reliability of the data.

[0217] The weights of the mapping vectors for the exponential growth effect increase exponentially with the amount of real-time interactive data. The weights grow slowly when the amount of data is small, but grow rapidly when the amount of data is large.

[0218] Linear decay means that the weight of the historical change baseline decreases linearly with the time interval, i.e. the time difference between the current and historical data. The weight is 1 - 0.01 × t, and the larger the interval, the lower the weight.

[0219] The fused feature vector is a comprehensive vector obtained by weighting and summing the feature vector of the historical baseline and the effect mapping vector according to the fusion weights. It contains historical patterns and real-time effect information. For example, after fusing the historical baseline vector {0.6, 0.7} with a weight of 0.3 and the effect mapping vector {0.8, 0.9} with a weight of 0.7, we get: {0.6×0.3+0.8×0.7, 0.7×0.3+0.9×0.7}={0.74, 0.84}.

[0220] Environmental variables are external environmental factors that affect the effectiveness of advertising within a preset time period, including holidays (such as National Day), weather (such as rainy days), and the intensity of competitor advertising (such as whether competitors are advertising at the same time).

[0221] The ad placement advantage prediction model is a machine learning model used to predict the ad placement advantage in each time unit. It takes fused features and environmental parameters as input and outputs a quantitative value. It adopts an LSTM model, takes fused vector and environmental parameters as input, and outputs a quantitative value of ad placement advantage of 0 to 1 for each hour.

[0222] The quantified advantage value is a quantitative indicator ranging from 0 to 1, output by the prediction model, reflecting the potential of the campaign effect in each time unit. The higher the value, the greater the campaign advantage in that time period.

[0223] Smoothing is a process of filtering the quantized values ​​of continuous time units to eliminate random fluctuations and highlight trends. For example, a 3-unit moving average can be applied to the quantized values ​​{0.6, 0.8, 0.7, 0.9, 0.8}, resulting in the smoothed values ​​{0.7, 0.77, 0.8, 0.85}.

[0224] Local peaks are the highest points in local time periods within the delivery advantage curve, reflecting the optimal time for short-term delivery. For example, the smoothed curve shows two local peaks of 0.8 and 0.85 in 12:00-13:00 and 18:00-19:00, respectively. Peaks are identified by finding the points in the curve where the derivative changes from positive to negative, and the corresponding time units and quantification values ​​are recorded.

[0225] The overall trend is the direction of change of the overall advantage curve, such as rising, falling, or fluctuating, reflecting the overall investment potential within a preset time period. The overall slope is calculated by linear fitting curves, with positive numbers indicating an upward trend and negative numbers indicating a downward trend.

[0226] In this embodiment, the delivery advantage curve is a curve plotted with time unit as the horizontal axis and delivery advantage quantification value as the vertical axis, which intuitively shows the changes in delivery advantage at different time periods.

[0227] The area under the curve is the area between the dominance curve and the horizontal axis, reflecting the total amount of dominance within a preset time period. Generally, the larger the area, the better the overall effect. The area under the curve is calculated using an integral algorithm.

[0228] Peak duration is the length of a continuous time unit containing a local peak, such as 2 hours. It reflects the sustainability of a high-dominance period. The duration of a continuous time unit where the quantified value is greater than 80% of the peak value is used as the peak duration.

[0229] Evaluation metrics are quantitative standards used to measure the final advantages of the campaign. Here, they refer to the area under the curve and the duration of the peak. For example, the evaluation metrics for ad placement A are: area under the curve 15.2 and peak duration 3 hours, which are better than ad placement B's 12.8 and 1 hour.

[0230] The beneficial effects of the above technical solution are: by breaking down historical baselines and analyzing virtual advertising effects in layers, and by combining historical and real-time data with dynamic weights, the final generated advertising advantage evaluation index can accurately reflect the advertising potential of each time period, effectively improving the efficiency and effectiveness of advertising.

[0231] This invention provides a method for optimizing an advertising delivery strategy, determining the ad prominence coefficient at each aligned time point, including:

[0232] Based on the quantitative value of the placement advantage of the time unit to which the alignment time point belongs, and combined with the placement interference coefficient of other ad positions at the corresponding time point, a basic prominence coefficient is constructed.

[0233] Introduce user activity weights corresponding to specific time points to dynamically adjust the basic prominence coefficient;

[0234] Extract the core visual features and content timeliness parameters of the ad placement at the corresponding time point, and generate feature enhancement factors;

[0235] The corrected salience coefficient and the feature enhancement factor are nonlinearly fused together. At the same time, the simulated interactive feedback of virtual ads at the corresponding time points is combined for secondary calibration to obtain the final ad placement salience coefficient.

[0236] Specifically, when the interference coefficient of an adjacent ad slot exceeds a first preset threshold, the feature enhancement factor weight of the current ad slot is automatically increased.

[0237] In this embodiment, the interference coefficient of other ad placements is the degree of interference of other ad placements to the current ad placement at the same time point, ranging from 0 to 1. The interference mainly comes from content similarity, such as both being women's clothing ads, and visual contrast, such as similar colors. The higher the value, the stronger the interference. The ad content relevance is calculated by a content similarity algorithm (such as cosine similarity), and the color difference is calculated by a visual contrast algorithm. The interference coefficient is obtained by weighted summation.

[0238] Basic prominence coefficient = Quantitative value of the advantage of the placement × (1 - Average interference coefficient).

[0239] In this embodiment, the user activity weight corresponding to the time point is the quantitative weight of the activity level of users around the ad position, such as the number of people staying and the movement speed, which is 0 to 1.2 at the aligned time point. The higher the activity, the greater the weight, which is used to enhance the prominence coefficient of high-activity periods.

[0240] In this embodiment, dynamic correction is the process of adjusting the basic prominence coefficient of user activity weight to make the corrected coefficient more in line with the actual user participation. The basic coefficient is combined with the activity weight through multiplication operation to output the corrected prominence coefficient.

[0241] In this embodiment, the core visual features are the visual attributes that are most likely to attract users' attention in the ad content displayed at the aligned time point, including the saturation of the main color tone, such as high saturation red, the occupancy of dynamic elements, such as 60% of the screen being animation, and the font size of key information, such as large titles.

[0242] In this embodiment, the content timeliness parameter is the time sensitivity of the advertising content from 0 to 1. For example, limited-time offers have high timeliness, while brand promotions with no time limit have low timeliness. The higher the value, the more the content needs to be delivered at the current time.

[0243] In this embodiment, the feature enhancement factor is an enhancement coefficient of 0.8 to 1.5 generated by integrating core visual features and content timeliness parameters. It is used to amplify the influence of the attractiveness of the advertising content itself on the salience coefficient. The more prominent the visual features and the higher the timeliness, the larger the factor. The quantitative score of the core visual features and the content timeliness parameters are weighted and summed, and then multiplied by the basic factor to obtain the final enhancement factor.

[0244] In this embodiment, the corrected prominence coefficient is the prominence coefficient adjusted after user activity weighting, which incorporates the impact of user activity on the campaign performance.

[0245] In this embodiment, nonlinear fusion combines the corrected salience coefficient with the feature enhancement factor through nonlinear functions such as exponential functions and multiplication operations, rather than simply adding them together, in order to amplify the synergistic effect of the two. For example, the combination of a high correction coefficient and a high enhancement factor will yield higher results. Here, y = a × (1 + b), where a is the correction coefficient and b is the enhancement factor.

[0246] In this embodiment, simulated interaction feedback is a quantitative feedback of 0 to 1 from simulated interaction data of virtual advertisement at the aligned time point, such as simulated click volume and simulated dwell time, used to verify the potential interaction effect of real advertisement.

[0247] In this embodiment, the secondary calibration is to readjust the nonlinear fusion result using the simulated interactive feedback of the virtual advertisement, so that the final coefficient is closer to the actual interaction potential. For example, if the nonlinear fusion result is 1.332 and the simulated interactive feedback of the virtual advertisement is 0.8, after secondary calibration, the result is 1.332 × 0.8 = 1.0656. The fusion result is multiplied by the simulated interactive feedback to achieve calibration and ensure that the coefficient reflects the real interaction possibility.

[0248] In this embodiment, the final ad placement prominence coefficient is a quantitative indicator that comprehensively reflects the prominence effect of ad placement at the aligned time point after secondary calibration. The higher the value, the easier it is for users to notice and respond to the ad placement at that time point.

[0249] In this embodiment, the interference coefficient of adjacent ad slots is the interference coefficient of ad slots that are physically adjacent to the current ad slot. The higher the value, the stronger the adjacent interference.

[0250] In this embodiment, the first preset threshold is a critical value, such as 0.6, for judging whether the adjacent interference is significant. If the value is exceeded, it is considered that the interference of the adjacent ad positions will seriously weaken the highlighting effect of the current ad position.

[0251] In this embodiment, the feature enhancement factor weight is automatically increased in nonlinear fusion when the adjacent interference coefficient exceeds a first preset threshold, such as from 50% to 70%, so as to counteract adjacent interference by enhancing its own content attractiveness.

[0252] The beneficial effects of the above technical solution are: by integrating the advantages of ad placement, interference coefficient, user activity, content characteristics and virtual advertising feedback, the ad placement highlight coefficient is calculated from multiple dimensions, and dynamically optimized for strong interference scenarios. The final coefficient can accurately reflect the outstanding effect of ad placement at each time point, providing a reliable basis for locking in the best placement node.

[0253] This invention provides a method for optimizing an advertising delivery strategy, comprising:

[0254] Based on the locked cut points, several advertising time segments are divided. Combining the advantages of advertising placements, user profile tags, and simulated conversion data of virtual ads within each time segment, a value assessment matrix for each segment is constructed.

[0255] Based on the value assessment matrix, the ad slots in each time segment are prioritized and the ad playback duration and frequency are dynamically adjusted.

[0256] Real-time public opinion data and competitor advertising dynamics are introduced as feedback factors. When the feedback factor triggers the second preset threshold, the backup advertising material library is automatically activated to replace advertising content with low conversion potential.

[0257] By comparing the ad exposure conversion rate, the increase in user dwell time, and the performance deviation rate between virtual and real ads before and after optimization, the threshold for determining the cut-off point and the granularity of time segment division are iteratively adjusted to form a self-optimizing ad placement strategy.

[0258] In this embodiment, the simulated conversion data of virtual advertising is the simulated conversion effect of virtual advertising in each time segment, such as the ratio of simulated clicks to purchases, which is used to predict the conversion potential of real advertising. For example, in the segment from 12:30 to 18:30, the simulated conversion data of virtual women's clothing advertising is: click to add to cart conversion rate 30%, add to cart to purchase conversion rate 20%. The simulated conversion logs for each time segment are extracted from the virtual advertising management system, and the conversion rate is calculated, such as simulated purchases / simulated clicks.

[0259] The value assessment matrix is ​​a matrix constructed with time segments as rows and assessment metrics as columns. Elements are quantitative values ​​from 0 to 1 for each segment in dimensions such as: campaign advantages, user profile matching degree, and virtual conversion data. It is used to comprehensively evaluate the value of each segment. For example, a value assessment matrix for three time segments:

[0260]

[0261] Priority ranking is based on the comprehensive score of the value assessment matrix, such as a weighted sum. The importance of the ad slots in each time segment is ranked, generally with level 1 being the highest and level 3 being the lowest. Resources are allocated to high-priority segments first. The matrix weighted score of each segment is calculated, and the segments are sorted in descending order of score and marked with priority level.

[0262] Dynamic adjustment is based on priority ranking results, adjusting the ad delivery parameters such as duration and frequency in real time for different time segments. Higher priority segments are allocated more resources. For example, the ad playback duration for a Level 1 segment from 12:30 to 18:30 is extended from 30 seconds to 45 seconds, and the frequency is increased from 4 times per hour to 6 times per hour; Level 3 segments remain at 20 seconds / 3 times. Through the ad delivery management system, the segment priority and playback parameters are linked, and automatic adjustment rules are set.

[0263] Ad playback duration is the continuous display time of a single ad on the ad slot, such as 30 seconds per ad. The longer the duration, the more information the user receives, but it may consume more resources.

[0264] Ad playback frequency refers to the number of times an ad is played per unit of time, such as 6 times per hour. The higher the frequency, the greater the exposure, but too high a frequency may cause user aversion.

[0265] Real-time public opinion data refers to real-time public opinion information related to the advertising content during the advertising period, such as brand reputation and product evaluation, including positive and negative sentiment tendencies, with a quantitative value of -1 to 1, and a negative value being negative.

[0266] Competitor campaign dynamics refers to the strategies employed by competitors in adjacent or similar ad slots within the same time period, such as ad content, playback frequency, and discounts. This reflects the intensity of market competition. The data is collected in real time through ad monitoring systems such as camera-based ad recognition and API integration with competitor campaign platforms to generate dynamic reports.

[0267] In this embodiment, the feedback factor is a quantitative indicator that integrates real-time public opinion data and competitor product dynamics. It is used to determine whether the product strategy needs to be adjusted. The higher the value, the more significant the change in the external environment. The feedback factor is generated by weighting and summing the absolute value of the public opinion data with the standardized values ​​of competitor product dynamics such as frequency / discount intensity.

[0268] The second preset threshold is the critical value of the feedback factor that triggers strategy adjustment. Exceeding this value indicates that changes in the external environment may significantly affect the campaign performance, and optimization needs to be initiated. For example, if the second preset threshold is set to 0.7, when the feedback factor reaches 0.8 > 0.7, the ad content replacement mechanism is triggered.

[0269] The backup ad creative library consists of pre-reserved alternative creatives that are related to the main ad theme but have different content, such as ads with different discounts or visual styles, to deal with unexpected situations.

[0270] Low conversion potential ad content refers to currently running ad content that has poor conversion performance based on feedback factors or data verification.

[0271] Ad impression conversion rate is the percentage of users who complete a target action, such as clicking or purchasing, after an ad is seen and exposed to the user. The formula is conversions / impressions, which reflects the actual effect of the ad.

[0272] The increase in user dwell time is the percentage increase in the average dwell time of users on the ad slot after optimization compared to before optimization. The formula is: (duration after optimization - duration before optimization) / duration before optimization × 100%, which reflects the increased attractiveness of the ad to users.

[0273] The deviation rate between virtual and real advertising is the proportion of the difference between the predicted effect of virtual advertising and the actual effect of real advertising.

[0274] Iterative adjustments involve continuously refining strategy parameters, such as cut-off thresholds and time segment granularity, based on performance metrics before and after optimization, such as conversion rate, increase in dwell time, and deviation rate, so that the strategy continuously approaches the optimal state.

[0275] The threshold for determining the cut point is a critical value, such as 0.7, used to determine whether the prominence coefficient meets the standard when screening cut points in the early stage. It can be increased or decreased after iterative adjustment, such as from 0.7 to 0.65. For example, the threshold parameter can be set in the strategy optimization module, and the threshold can be lowered by 5% when the conversion rate increase is greater than 20%.

[0276] The granularity of time segment division is the size of the time interval between time segments. For example, if the original 1 hour / segment is adjusted to 30 minutes / segment, the finer the granularity, the more accurate the strategy, but the higher the complexity. Adjustments are made based on the variance of the effect under different granularities. The larger the variance, the finer the granularity is required. The granularity value can be modified through the parameter configuration module.

[0277] Self-optimizing advertising strategies automatically adapt to changes in the environment, such as public opinion, competitors, and performance feedback, through iterative adjustments, and can be dynamically optimized without manual intervention.

[0278] The beneficial effects of the above technical solution are: by dividing time segments by cutting points and constructing a value assessment matrix, combining priority with dynamic resource allocation, and introducing real-time feedback to adjust content, the strategy parameters are iteratively optimized through effect comparison, resulting in a self-optimizing delivery strategy that significantly improves ad exposure and conversion efficiency and user attention, while reducing resource waste.

[0279] This invention provides an optimization system for advertising delivery strategies, such as... Figure 2 As shown, it includes:

[0280] The matrix construction module is used to obtain multi-dimensional attribute information of each ad slot on the target display screen from the record database and perform quantitative processing to construct an attribute representation matrix. The multi-dimensional attribute information is related to pre-ad active playback information, pre-ad passive playback information, traffic volume based on each passive playback, added virtual ads, and the playback status of virtual ads.

[0281] The feature acquisition module is used to extract the representation vector of each attribute representation matrix and, depending on the original distribution of the ad slots on the target display screen, obtain the multidimensional features of each ad slot. The multidimensional features include: main representation features and auxiliary representation features of the remaining ad slots for the corresponding ad slots.

[0282] The advantage determination module is used to match historical change baselines that match each multi-dimensional feature from the historical delivery database, and based on the historical change baselines, combined with the added virtual advertisements, predict the delivery advantage of the corresponding advertisement position within a preset time period.

[0283] The segmentation optimization module is used to perform time alignment processing based on the advantages of all ad placements, determine the ad placement prominence coefficient at each aligned time point, and lock the time points with prominence coefficients greater than preset coefficients as segmentation points to optimize the ad placement strategy.

[0284] This invention provides a platform for executing an optimization method for any of the described advertising delivery strategies.

[0285] The beneficial effects of the above technical solution are: by comprehensively quantifying the characteristics of ad placements with multi-dimensional attributes, mining the correlation between ad placements, integrating historical data with the advantages of virtual advertising prediction, and accurately optimizing based on time nodes, the targeting and effectiveness of advertising can be effectively improved, resource waste can be reduced, and the anti-interference ability and adaptability of the strategy can be enhanced through virtual advertising and interference processing.

[0286] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An advertisement delivery strategy optimization method, characterized by, The method comprises the following steps: Step 1: obtaining multi-dimensional attribute information of each advertising position on a target display screen from a record database and performing quantitative processing to construct an attribute representation matrix, wherein the multi-dimensional attribute information is related to pre-ad active playing information, pre-ad passive playing information, traffic size based on each passive playing, added virtual advertisement and playing situation of the virtual advertisement; Step 2: extracting a representation vector of each attribute representation matrix, and obtaining multi-dimensional features of each advertising position in dependence on original distribution of the advertising positions on the target display screen, wherein the multi-dimensional features comprise main representation features and auxiliary representation features of the remaining advertising positions to the corresponding advertising position; Step 3: matching a historical change baseline matched with each multi-dimensional feature from a historical delivery database, and predicting a delivery advantage of the corresponding advertising position in a preset time period in dependence on the historical change baseline and in combination with the added virtual advertisement; Step 4: performing time alignment processing according to the delivery advantages of all the advertising positions, determining an advertising delivery highlighting coefficient of each alignment time point, and locking a time point with a highlighting coefficient greater than a preset coefficient as a cutting point to optimize an advertising delivery strategy; Wherein, obtaining multi-dimensional attribute information of each advertising position on a target display screen from a record database and performing quantitative processing to construct an attribute representation matrix comprises: inputting the multi-dimensional attribute information into a quantitative processing model matched with the target display screen to obtain a multi-dimensional output vector of the corresponding advertising position, simultaneously, obtaining user behavior dynamic features of each advertising position and spatial correlation parameters of adjacent advertising positions as extended dimensions of the multi-dimensional attribute information, and quantitatively obtaining an extended vector; determining a mapping priority according to a mapping relationship between each extended dimension and each dimension type in the multi-dimensional attribute information, and generating a mapping path; constructing a topology network according to the mapping path and in combination with an advertising change set of each dimension type; performing element adjustment on the topology network to determine a reserved parameter of the extended dimension and an influence factor of each reserved parameter on the corresponding original element under the mapping relationship, and obtaining a new vector; mapping all the new vectors to a high-dimensional feature space to generate an attribute representation matrix.

2. The method of claim 1, wherein, Obtaining an added virtual advertisement comprises: obtaining generation materials, visual effects and delivery environment of each advertising position based on the target display screen to construct delivery bias of each advertising position; determining a first sequence of historical brushing attributes and a second sequence of delivery bias of each advertising position; extracting feature parameters of the first sequence and the second sequence, and determining a connection relationship between the feature parameters, and adding an initial advertisement to the corresponding advertising position; performing single-frame splitting on the initial advertisement to construct a brushing elimination directed graph of each split frame, and obtaining an elimination probability; dividing the initial advertisement by using an interactive window, and determining elimination features of each divided sub-advertisement in combination with the elimination probability; performing linear detection on all the elimination features, extracting optimal features, and optimizing the initial advertisement to obtain the virtual advertisement in combination with a global coverage ratio and a trapping intensity of the optimal features.

3. The method of claim 1, wherein, Obtaining multi-dimensional features of each advertising position in dependence on original distribution of the advertising positions on the target display screen comprises: Nonlinear dimensionality reduction is performed on the attribute representation matrix to extract deep principal representation vectors containing high-order feature interactions; An ad space correlation graph is constructed based on the physical coordinate distribution of the target display screen, wherein the nodes are ad slots and the edge weights are the visual attention transfer probability and the spatiotemporal co-occurrence frequency of adjacent ad slots; Feature propagation is performed on the spatial correlation graph using a graph neural network to extract node noise features and path segment noise features between nodes in each propagation path during feature propagation, construct a noise sequence, and extract local and global significant noise factors; Noise elimination processing is performed on the feature propagation process based on all local and global significant noise factors to generate a spatial correlation feature vector for each ad slot as an auxiliary representation feature, wherein a distance decay coefficient is introduced in the feature propagation process to linearly decay the feature influence of adjacent ad slots as the physical distance increases; Temporal dynamic feature vectors are extracted as supplementary auxiliary representation features by combining the feature fluctuation sequence of the ad slots in the preset historical period. Based on the deep principal representation vector, the spatial correlation feature vector, and the temporal dynamic feature vector, a multi-dimensional feature of the corresponding ad slot is generated.

4. The method of claim 1, wherein, Based on the historical change baseline and combined with the added virtual ad, the prediction of the delivery advantage of the corresponding ad slot in the preset time period includes: The historical change baseline is decomposed into multiple sub-baseline sequences according to the time granularity, wherein each sub-baseline sequence corresponds to a time unit in the preset time period, and the trend feature parameters and fluctuation thresholds of each sub-baseline sequence are extracted; The play data of the added virtual ad is analyzed hierarchically to obtain the interaction conversion rate and content memory retention of the virtual ad in different user portrait groups, and a two-dimensional portrait matrix of the virtual ad is constructed; Each dimension in the two-dimensional portrait matrix is arranged in descending order according to the element value, and a serial number group of each user remaining after the last user in the sorted order is constructed; A difference matrix is constructed based on the difference between the adjacent two rows of the matrix sorted in descending order, each column result in the difference matrix is clustered and classified, and the mode range of each column result is determined, the minimum value in the mode range is used to replace the element values before the mode range, and the maximum value in the mode range is used to replace the element values after the mode range to obtain a new two-dimensional matrix, and a two-dimensional array of each remaining user is obtained; An effect mapping vector is obtained based on the two-dimensional array of each remaining user, the serial number group, and the original array; The fusion weight of the historical change baseline and the effect mapping vector is dynamically adjusted based on a time decay function, wherein the weight of the effect mapping vector increases exponentially with the increase of real-time interaction data, and the weight of the historical change baseline decreases linearly with the extension of time interval; The fused feature vector and the environmental variable parameters in the preset time period are input into the delivery advantage prediction model to output the delivery advantage quantitative value of each time unit; The delivery advantage quantitative values of consecutive time units are smoothed to generate a delivery advantage curve containing local peaks and global trends, and the area under the curve and the peak duration are used as evaluation indicators of the final delivery advantage.

5. The method of claim 4, wherein, The advertisement exposure prominence coefficient of each alignment time point is determined, including: Based on the exposure advantage quantitative value of the time unit to which the alignment time point belongs, combined with the exposure interference coefficient of other advertisement positions at the corresponding time point, a basic prominence coefficient is constructed; The user activity weight corresponding to the time point is introduced to dynamically correct the basic prominence coefficient; The core visual features and content timeliness parameters of the advertisement position at the corresponding time point are extracted to generate a feature enhancement factor; The corrected prominence coefficient and the feature enhancement factor are nonlinearly fused, and at the same time, the final advertisement exposure prominence coefficient is obtained by secondary calibration combined with the simulated interaction feedback of the virtual advertisement at the corresponding time point; Wherein, when the interference coefficient of the adjacent advertisement position exceeds the first preset threshold, the feature enhancement factor weight of the current advertisement position is automatically increased.

6. The method of claim 1, wherein, The advertisement exposure strategy is optimized, including: Based on the locked cutting point, a plurality of advertisement exposure time segments are divided, and a value evaluation matrix of each segment is constructed combined with the exposure advantage, user portrait label and simulated conversion data of the virtual advertisement in each time segment; Based on the value evaluation matrix, the advertisement positions in each time segment are prioritized, and the advertisement playing time and frequency are dynamically adjusted; Real-time public opinion data and competitor exposure dynamics are introduced as feedback factors, and when the feedback factors trigger the second preset threshold, the standby advertisement material library is automatically activated to replace the advertisement content with low conversion potential; The advertisement exposure conversion rate, user stay time increase and effect deviation rate of virtual advertisement and real advertisement before and after optimization are compared, and the judgment threshold of the cutting point and the division granularity of the time segment are iteratively adjusted to form a self-optimizing advertisement exposure strategy.

7. An advertisement delivery strategy optimization system, characterized by, Including: The matrix construction module is used to obtain the multi-dimensional attribute information of each advertisement position on the target display screen from the record database and perform quantization processing to construct an attribute representation matrix, wherein the multi-dimensional attribute information is related to pre-advertiser active playing information, pre-advertiser passive playing information, traffic size based on each passive playing, added virtual advertisement and playing situation of the virtual advertisement; The feature acquisition module is used to extract the representation vector of each attribute representation matrix, and to obtain the multi-dimensional features of each advertisement position depending on the original distribution of the advertisement positions on the target display screen, wherein the multi-dimensional features include main representation features and auxiliary representation features of the corresponding advertisement positions by the remaining advertisement positions; The advantage determination module is used to match the historical change baseline matched with each multi-dimensional feature from the historical exposure database, and to predict the exposure advantage of the corresponding advertisement position in a preset time period depending on the historical change baseline and combined with the added virtual advertisement; The cutting optimization module is used to perform time alignment processing according to the exposure advantages of all advertisement positions, to determine the advertisement exposure prominence coefficient of each alignment time point, and to lock the time points with prominence coefficients greater than a preset coefficient as cutting points to optimize the advertisement exposure strategy; Wherein, the matrix construction module is used to: inputting the multi-dimensional attribute information into a quantization processing model matched with the target display screen to obtain a multi-dimensional output vector of the corresponding advertisement position, and meanwhile, obtaining user behavior dynamic characteristics of each advertisement position and spatial correlation parameters of adjacent advertisement positions as extended dimensions of the multi-dimensional attribute information, and obtaining an extended vector by quantization; determining a mapping priority according to a mapping relationship between each extended dimension and each dimension type in the multi-dimensional attribute information, and generating a mapping path; constructing a topological network according to the mapping path and in combination with an advertisement change set of each dimension type; performing analysis on the topological network to determine a reserved parameter of the extended dimension and an influence factor of each reserved parameter on a corresponding original element under the mapping relationship, and performing element adjustment to obtain a new vector; mapping all new vectors to a high-dimensional feature space to generate an attribute representation matrix.

8. A platform, characterized by An optimization method for performing the advertisement delivery strategy of any one of claims 1-6.

Images