AI large model driven short video advertisement putting optimization system

Abstract

The application relates to the technical field of machine learning, in particular to an AI large model driven short video advertisement putting optimization system, which comprises a node feature extraction module, a turning point detection module, a resource scheduling module and an interest trend analysis module.In the application, through deep semantic vectorization of multidimensional data, dynamic capture of user behavior and content interest, perception of key turning points of node influence, fine-grained classification according to user interest change trend, dynamic adjustment of advertisement pushing parameters, real-time weight optimization of resource allocation strategy, high adaptation of advertisement pushing content to user interest, continuous improvement of advertisement putting effect, improvement of advertisement resource utilization rate, positive improvement of user experience, whole process cooperation from multidimensional data collection to putting strategy optimization, quick response to user interest deviation, enhancement of advertisement matching and interaction performance, and yield improvement and data value release for short video advertisement business are realized.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a short video advertising optimization system driven by AI large models. Background Technology

[0002] Machine learning technology involves using data-driven methods to automatically model and predict complex problems. This field includes core aspects such as feature engineering, model training, model evaluation, and optimization. It is widely applied in various practical scenarios, including image recognition, speech recognition, natural language processing, recommendation systems, and advertising. The entire technology is based on large-scale data, training models to uncover potential patterns within the data and applying them to real-world business decisions. Traditional short video advertising optimization systems, for example, optimize advertising performance on short video platforms based on multi-dimensional data such as user behavior history, interests, preferences, content characteristics, and advertising materials. Common practices include using machine learning models to score the match between users and ads and ranking them according to the scores, and training models using historical ad click or conversion data to predict the performance of different advertising materials or strategies among target user groups, thereby achieving reasonable allocation of advertising resources and dynamic adjustment of content.

[0003] Existing technologies often rely on historical behavior and interest preferences for model training and matching during ad delivery. Due to a lack of sensitivity to changes in user interests, resource allocation is somewhat lagging. In scenarios with rapid fluctuations in interests or changes in content trends, ad delivery struggles to respond promptly to user focus, resulting in homogenized ad pushes, affecting personalized recommendation effectiveness, and causing inefficient allocation of some ad resources. User experience declines due to mismatch between ads and actual interests. When trending topics appear on short video platforms, the system fails to capture interest shifts in a timely manner, causing ad push content to become disconnected from real-time user needs, impacting the value of advertisers' campaigns and platform revenue. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-driven large-scale model-based short video advertising optimization system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An AI-driven short video advertising optimization system includes:

[0006] The node feature extraction module is used to collect user behavior sequence parameters, content tags, and interest semantic parameters to obtain multi-dimensional node parameter data, input it into the BERT model for feature vectorization, call the node feature encoding results, and pass the node feature encoding results to the inflection point detection module.

[0007] The inflection point detection module is used to obtain a node behavior feature sequence based on the node feature encoding result, call the node behavior feature sequence to input the temporal attention mechanism algorithm, detect and identify the node influence inflection point, and pass the node influence inflection point to the resource scheduling module.

[0008] The resource scheduling module is used to obtain the node influence inflection point, adjust the weight of the advertising resource parameter set using the multi-armed slot machine algorithm, and transmit the resource adjustment instruction to the interest trend analysis module.

[0009] The interest trend analysis module is used to obtain the resource adjustment instruction, use an interest trend clustering algorithm to classify the node interest trends, and adjust the subsequent advertising push parameters based on the node category and the resource adjustment instruction.

[0010] As a further aspect of the present invention, the node feature encoding result includes user interest vector, content relevance vector, and behavior preference vector; the node influence inflection point specifically refers to node behavior change point, user decision threshold point, and influence fluctuation point; the resource adjustment instruction includes advertising resource allocation weight, placement time adjustment, and placement priority parameter; and the node interest trend category specifically refers to interest growth category, interest decay category, and interest fluctuation category.

[0011] As a further aspect of the present invention, the node feature extraction module includes:

[0012] The behavior and interest data collection submodule collects user behavior sequence parameters, content tags, and interest semantic parameters. Based on user interaction logs, it calculates the number of clicks on multiple content tags and the average user dwell time. Combining the frequency of user feedback on differentiated interest semantics, it integrates multiple statistical data to generate raw user behavior indicators.

[0013] The multi-dimensional node parameter construction submodule calls the original user behavior indicators, filters the user behavior sequence parameters based on the preset interaction activity benchmark value, removes sequences with interaction frequency lower than the benchmark value, and integrates the filtered behavior sequence parameters with content tags and interest semantic parameters in a structured manner to establish multi-dimensional node parameter data.

[0014] The node feature vectorization submodule inputs the multi-dimensional node parameter data into the BERT model, uses the self-attention mechanism inside the model to calculate the correlation score between multiple dimension parameters, captures the deep contextual relationships in the parameter sequence through a multi-layer encoder structure, maps discrete node parameters to a continuous vector space, and obtains the node feature encoding results.

[0015] As a further aspect of the present invention, the inflection point detection module includes:

[0016] The behavior feature serialization submodule extracts the timestamp information from each encoding result based on the node feature encoding result, reorders all node feature encodings according to the order of the timestamps, and concatenates the sorted encodings into a time-continuous data stream to generate a node behavior feature sequence.

[0017] The temporal attention weight calculation submodule calls the node behavior feature sequence to input into the temporal attention mechanism algorithm, calculates the query vector and key vector for the feature vector of each node in the sequence, obtains the attention score through vector dot product operation, and then uses the Softmax function to normalize all scores to obtain the node temporal attention score.

[0018] The influence turning point determination submodule performs a weighted summation of the node behavior feature sequence based on the node temporal attention score, calculates the context representation vector of each time node, monitors the change amplitude of the cosine similarity of the context representation vectors between adjacent nodes, compares the change amplitude with the set influence change threshold, identifies nodes whose change amplitude exceeds the threshold, and obtains the node influence turning point.

[0019] As a further aspect of the present invention, the resource scheduling module includes:

[0020] The inflection point resource mapping submodule obtains the node influence inflection point, collects the set of advertising placement resource parameters, analyzes the user group profile associated with the inflection point and the target audience tags of multiple advertising resources, calculates the overlap between the two, and filters the advertising resources associated with the inflection point according to the preset audience matching threshold to establish a list of candidate placement resources.

[0021] The slot machine revenue assessment submodule calls the list of candidate deployment resources, treats each resource item in the list as an independent arm of the multi-arm slot machine algorithm, and calculates the upper confidence interval of each arm based on the historical conversion rate and deployment cost data of multiple resource items, and obtains the expected revenue value of the resource.

[0022] The resource weight dynamic adjustment submodule sorts the resources in the candidate placement resource list according to the expected revenue value of multiple advertising resources, uses the sorting result as the basis for weight allocation, recalculates the allocation ratio of multiple resources in the placement budget through a normalization function, and integrates the resource identifier with the parameters of the new allocation ratio to generate a resource adjustment instruction.

[0023] As a further aspect of the present invention, the interest trend analysis module includes:

[0024] The interest feature extraction submodule obtains the resource adjustment instruction, collects the user interaction logs of the advertising resources in the instruction, counts the interaction frequency and duration data between each node and the associated advertisement, weights and fuses the statistical data, calculates the interest intensity of each node, and establishes a node interest feature vector.

[0025] The trend clustering analysis submodule calls the node interest feature vector, applies the interest trend clustering algorithm, uses Euclidean distance as the measure of similarity between vectors, it iteratively calculates and assigns the feature vector to the nearest cluster center, and continuously updates the position of the cluster center until convergence, dividing the nodes in each cluster into a class, and obtaining the node interest trend category.

[0026] The push parameter adjustment submodule, based on the node interest trend category and the resource adjustment instruction, adjusts the direction and magnitude of the interest trend represented by each category according to the weight in the instruction, quantitatively modifies the delivery frequency and target audience range in the ad push parameters, generates an instruction set of updated parameter values, and obtains subsequent ad push parameters.

[0027] As a further aspect of the present invention, the process of calculating the interest intensity of each node specifically involves first normalizing the collected interaction frequency and duration data, and then using the formula... Perform calculations;

[0028] in, Represents the interest intensity of node n. This represents the normalized interaction frequency between node n and associated ads. The normalized interaction duration between node n and associated ads. Weighting coefficients representing normalized interaction frequencies. The weight coefficient represents the normalized interaction duration, where n is the node index, and satisfies... .

[0029] As a further aspect of the present invention, the slot machine revenue evaluation submodule specifically includes:

[0030] For the i-th ad resource in the candidate ad resource list, collect its total historical ad placement count after t rounds of placement. Total number of historical ad placements across all ad resources And calculate its historical average conversion rate. ;

[0031] Call the historical average conversion rate The total number of historical deployments Compared with the total number of historical ad deliveries for all the aforementioned ad resources Through the preset upper confidence limit formula Calculations are performed to obtain the expected revenue value of each advertising resource in the list of candidate advertising resources;

[0032] in, This represents the expected return on the i-th ad resource in the t-th round of deployment. This represents the index of the ad resource in the list of potential ad delivery resources. This represents the round of deployment.

[0033] As a further aspect of the present invention, the influence transition determination submodule specifically includes:

[0034] Collect the cosine similarity variation data of the context representation vectors between all adjacent nodes within a historical time window, and calculate the mean of the cosine similarity variation data. with standard deviation ;

[0035] Call the mean The standard deviation and preset risk coefficient Through the dynamic threshold formula: The threshold for the change in influence is calculated.

[0036] in, The threshold representing changes in influence. This is a preset risk coefficient constant.

[0037] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0038] In this invention, through deep semantic vectorization of multidimensional data and dynamic capture of user behavior and content interests, key turning points in node influence can be accurately perceived. Fine-grained classification of user interest change trends allows for dynamic adjustment of ad push parameters, effectively achieving real-time weight optimization of resource allocation strategies. This results in a high degree of alignment between ad push content and user interests, thereby driving continuous improvement in ad performance, significantly improving ad resource utilization, and positively enhancing user experience. The entire process achieves efficient collaboration from multidimensional data collection to ad delivery strategy optimization, enabling rapid response to user interest shifts, enhancing ad matching and interactive performance, and ultimately bringing increased revenue and data value release to the short video advertising business. Attached Figure Description

[0039] Figure 1 This is a flowchart of the AI-driven short video ad delivery optimization system of the present invention;

[0040] Figure 2 This is a flowchart of the node feature extraction module of the present invention;

[0041] Figure 3 This is a flowchart of the inflection point detection module of the present invention;

[0042] Figure 4 This is a flowchart of the resource scheduling module of the present invention;

[0043] Figure 5 This is a flowchart of the interest trend analysis module of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0045] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0046] Please see Figure 1 and Figure 2 This invention provides a technical solution: an AI-driven short video advertising optimization system includes:

[0047] The node feature extraction module is used to collect user behavior sequence parameters, content tags, and interest semantic parameters to obtain multi-dimensional node parameter data, input it into the BERT model for feature vectorization, call the node feature encoding results, and pass the node feature encoding results to the inflection point detection module.

[0048] The node feature encoding results include user interest vectors, content relevance vectors, and behavioral preference vectors;

[0049] The node feature extraction module includes:

[0050] The behavior and interest data collection submodule collects user behavior sequence parameters, content tags, and interest semantic parameters. Based on user interaction logs, it calculates the number of clicks on multiple content tags and the average user dwell time. Combining the frequency of user feedback on differentiated interest semantics, it integrates multiple statistical data to generate raw user behavior indicators.

[0051] The node feature extraction module receives user behavior sequence parameters, content tags, and interest semantic parameters from the front-end data acquisition system. Specifically, the behavior and interest data acquisition submodule first collects user interaction logs within a time window (e.g., the past 24 hours). This log records a series of behaviors for user ID U001, with parameters including user ID, video ID, viewing timestamp, viewing duration (seconds), whether the user liked the video (1 yes / 0 no), whether they commented on it (1 yes / 0 no), whether they shared it (1 yes / 0 no), and viewing completion rate (%). Simultaneously, the system collects video-related content tags, such as #TechReview, #OutdoorCamping, #FitnessTutorial, and the user's selected interest semantics on specific interactive interfaces, such as in-depth understanding and entertainment.

[0052] During execution, the submodule calculates user U001's behavior under the #TechReview tag. In the past 24 hours, user U001 clicked on videos tagged with #TechReview 8 times, totaling 8 clicks. The total viewing time for these 8 videos was 450 seconds, therefore the average user dwell time was [missing value]. Seconds. Regarding semantic interest, after watching three of the #TechReview videos, user U001 clicked the "Learn More" feedback button, with a feedback frequency of [missing information]. These calculated data, along with statistical results from other tags and semantics, are integrated into the raw behavioral metrics for user U001, specifically in the format {User ID:U001, Metrics:[{Tag:#Tech Review, Click Count: 8, Average Dwell Time: 56.25, Semantic Feedback:{In-depth Understanding: 0.375}},{Tag:#Outdoor Camping, Click Count: 2, Average Dwell Time: 15.5, Semantic Feedback:{Entertainment: 1.0}}]}.

[0053] The multidimensional node parameter construction submodule calls the user's original behavior indicators, filters the user behavior sequence parameters based on the preset interaction activity benchmark value, removes sequences with an interaction frequency lower than the benchmark value, and integrates the filtered behavior sequence parameters with content tags and interest semantic parameters in a structured way to establish multidimensional node parameter data.

[0054] The multi-dimensional node parameter construction submodule calls the aforementioned user raw behavior metrics. This submodule configures an interaction activity benchmark value, which is set based on statistical analysis of the behavior data of all users on the platform over the past 30 days. The specific experimental process is as follows: The total number of interactions (likes, comments, and shares) of 1 million random sample users over 30 days is collected and sorted. The interaction frequency distribution exhibits a long-tail effect; to filter out low-frequency and invalid users, the interaction frequency at the 20th percentile is selected as the benchmark value. In this sample data, the 20th percentile has 5 interactions per week. Therefore, the interaction activity benchmark value is set to 5 times / week. User U001 had a total of 12 interactions in the past week (including likes and comments on #TechReview and other videos). Therefore, the user's behavioral sequence parameters are deemed valid and retained. If another user, U002, has a total of 3 interactions, which is below the baseline, their behavioral sequence will be removed. After filtering, the valid behavioral sequence parameters are structurally integrated with the corresponding content tags and interest semantic parameters. For example, user U001's data points regarding #TechReview are integrated into a structured record: {Timestamp: 20250728T103000Z, User ID: U001, Behavior: Click, Viewing Duration: 68, Likes: 1, Comments: 0, Shares: 0, Completion Rate: 0.9, Tags: #TechReview, Semantics: In-depth Understanding}. Multiple such record sets constitute multi-dimensional node parameter data.

[0055] The node feature vectorization submodule inputs multi-dimensional node parameter data into the BERT model, uses the model's internal self-attention mechanism to calculate the correlation scores between multiple dimension parameters, captures deep contextual relationships in the parameter sequence through a multi-layer encoder structure, maps discrete node parameters to a continuous vector space, and obtains node feature encoding results.

[0056] The node feature vectorization submodule receives structured, multi-dimensional node parameter data. This discrete, multi-dimensional parameter data, such as behavior (click) and tag (#techreview), is first mapped to a pre-defined dictionary and converted into unique numeric IDs. These ID sequences are then input into a 768-dimensional BERT model. Inside the model, a self-attention mechanism calculates the association score between each parameter ID in the sequence and all other parameter IDs. For example, the association score between "view duration: 68" and "completion rate: 0.9" and the semantic meaning "deep understanding" will be higher than its score with the tag "#outdoorcamping". This process is repeated across the model's 12 encoder layers, each layer capturing deeper contextual connections based on the previous layer. For example, lower layers might only focus on the direct association between "likes" and "#techreview", while higher layers can capture the abstract concept of "strong professional interest" jointly pointed to by "high view completion rate", "deep understanding" semantics, and the "#techreview" tag. Through these 12 layers of encoding, the discrete input parameter sequence is ultimately mapped into a 768-dimensional continuous vector space. This vector is the node feature encoding result, which includes user interest vector (e.g., high values ​​of certain dimensions in the vector correspond to scientific and technological interests), content relevance vector (reflecting the relationship between video content), and behavioral preference vector (encoding the user's interaction habits). This encoding result [-0.21, 0.89, 0.15, ..., -0.55] is then passed to the inflection point detection module.

[0057] Please see Figure 1 and Figure 3 The inflection point detection module is used to obtain the node behavior feature sequence based on the node feature encoding result, call the node behavior feature sequence to input the temporal attention mechanism algorithm, detect and identify the node influence inflection point, and pass the node influence inflection point to the resource scheduling module.

[0058] The turning points of node influence are specifically the points of change in node behavior, the user decision threshold points, and the points of fluctuation in influence.

[0059] The inflection point detection module includes:

[0060] The behavior feature serialization submodule extracts the timestamp information from each encoding result based on the node feature encoding results, reorders all node feature encodings according to the order of the timestamps, and concatenates the sorted encodings into a time-continuous data stream to generate a node behavior feature sequence.

[0061] The behavior feature serialization submodule first processes these encoding results. It extracts the timestamp information from each encoding result, for example, T1: 20250728T103000Z, T2: 20250728T103500Z, T3: 20250728T104000Z. Based on the chronological order of these timestamps, the submodule reorders all feature encoding vectors for user U001. After sorting, these independent vectors are concatenated into a temporally continuous data stream, i.e., [vector_T1, vector_T2, vector_T3,...], which constitutes the node behavior feature sequence.

[0062] The temporal attention weight calculation submodule calls the node behavior feature sequence as input to the temporal attention mechanism algorithm, calculates the query vector and key vector for the feature vector of each node in the sequence, obtains the attention score through vector dot product operation, and then uses the Softmax function to normalize all scores to obtain the node temporal attention score.

[0063] The temporal attention weight calculation submodule receives the sequence of behavioral features of the node and inputs it into the temporal attention mechanism algorithm. The algorithm calculates a query vector for the feature vector (e.g., vector_T2) of each node in the sequence. and a key vector This is achieved by multiplying the vector _T2 by two different weight matrices learned during the training phase. and Obtained. Subsequently. The key vectors of all time points in the sequence (including T1, T2, T3, etc.) , , Perform a vector dot product operation to obtain the raw attention score. For example, the score... , , Assume the original scores are [1.5, 3.2, 0.8]. These scores are then normalized using the Softmax function. The calculation process is as follows: ; ; The normalized score [0.143, 0.785, 0.072] is the node temporal attention score at time T2, which indicates how much attention the system should give to the features of time points T1, T2, and T3 at time T2.

[0064] The influence turning point determination submodule performs weighted summation on the node behavior feature sequence based on the node temporal attention score, calculates the context representation vector of each time node, monitors the change amplitude of the cosine similarity of the context representation vectors between adjacent nodes, compares the change amplitude with the set influence change threshold, identifies nodes whose change amplitude exceeds the threshold, and obtains the node influence turning point.

[0065] The influence turning point determination submodule specifically includes:

[0066] Collect data on the cosine similarity variation of the context representation vectors between all adjacent nodes within a historical time window, and calculate the mean of the cosine similarity variation data. with standard deviation ;

[0067] Call the mean Standard deviation and preset risk coefficient Through the dynamic threshold formula: Calculations were performed to obtain the threshold for changes in influence;

[0068] in, The threshold representing changes in influence. This is a preset risk coefficient constant.

[0069] The influence turning point determination submodule calculates the context representation vector for each time node by weighted summing of the node behavior feature sequences based on the aforementioned node temporal attention scores. For example, the context representation of T2... The submodule continuously monitors the context representation vectors between adjacent nodes (e.g., T2 and T3). and The variation range of the cosine similarity. The calculated cosine similarity result is between -1 and 1, and the variation range is... The magnitude of this change is related to a dynamically set threshold for changes in influence. The comparison is performed, and if the magnitude exceeds the threshold, the corresponding node is identified as a turning point in the node's influence.

[0070] The threshold for this influence change Through dynamic threshold formula The calculations are as follows. A detailed explanation of each parameter in the formula is provided below: This represents the threshold value for the change in influence ultimately used for judgment; It is the mean of the cosine similarity change data of the context representation vectors between all adjacent nodes within a historical time window; It is the standard deviation of the same data sample; This is a preset risk coefficient constant used to adjust the sensitivity of the threshold. The calculation logic of this formula lies in establishing a dynamic benchmark that adaptively adjusts as the stability of the user's recent behavior changes. Mean It represents the average level of changes in user behavior, while the standard deviation... This quantifies the degree of fluctuation in this change. This is achieved by adding a factor of the standard deviation (from...). (Control), thresholds can effectively identify statistically significant abnormal fluctuations, rather than simple absolute value changes.

[0071] To determine the parameter values ​​in the formula, the system first collects data on the cosine similarity changes of the context representation vectors between all adjacent nodes of user U001 over the past hour (historical time window), resulting in an array: [0.021, 0.025, 0.020, 0.280, 0.031, 0.028, 0.022, 0.025]. Based on this data, the mean is calculated. and standard deviation Risk coefficient The settings were validated through offline experiments. The experiments used historical behavioral data from 10,000 users, including 500 interest inflection points manually labeled by operations experts. Different... Backtesting was performed using values ​​(from 1.0 to 3.0, in steps of 0.5), and the precision and recall of the detection results were evaluated.

[0072] Table 1 Risk Coefficient Experimental Data Table

[0073]

[0074] As shown in Table 1, when At that time, the F1 score reached its highest value of 0.874, achieving a good balance between precision and recall. Therefore, the system presets a risk coefficient. The value is 1.5. Substitute the above parameter value into the formula to calculate: The calculated threshold for the change in influence is 0.1912. The advantage of this formula lies in the introduction of the mean. and standard deviation This allows the threshold to adapt to the volatility of each user's behavior, rather than using a rigid threshold that is fixed for all users; at the same time, the risk coefficient... The introduction of this feature provides an interface for adjusting strategies at the business level. In actual monitoring, the change in a certain node was 0.280, due to... This node was identified as a turning point in node influence. The result indicates that the user's behavior at this point in time exceeded its normal fluctuation range, constituting a statistically significant turning point. This turning point information, specifically tagged as {User ID: U001, Timestamp: 20250728T104500Z, Turning Point Type: Node Behavior Change Point}, was transmitted to the resource scheduling module.

[0075] Please see Figure 1 and Figure 4 The resource scheduling module is used to obtain the turning point of node influence, adjust the weight of the advertising resource parameter set using the multi-armed slot machine algorithm, and transmit the resource adjustment instructions to the interest trend analysis module.

[0076] Resource adjustment instructions include advertising resource allocation weights, ad placement time adjustments, and ad placement priority parameters;

[0077] The resource scheduling module includes:

[0078] The Turning Point Resource Mapping submodule obtains the node influence turning point, collects the set of advertising resource parameters, analyzes the user group profile associated with the turning point and the target audience tags of multiple advertising resources, calculates the overlap between the two, and filters the advertising resources associated with the turning point according to the preset audience matching threshold to establish a list of candidate advertising resources.

[0079] The inflection point resource mapping submodule first obtains the inflection point information and collects the set of currently available advertising resource parameters. This set contains three advertising resources: Ad_A (new smartwatch, target audience tags: technology, sports, efficiency), Ad_B (branded coffee, target audience tags: lifestyle, leisure, white-collar workers), and Ad_C (online programming course, target audience tags: technology, education, workplace). Simultaneously, the submodule analyzes the user profile of user U001 associated with the inflection point; their latest user interest vector indicates that their core interest tags are {technology, workplace, efficiency}. The submodule calculates the overlap between the user profile tags and the target audience tags of each advertising resource, using the Jaccard similarity coefficient for calculation.

[0080] The overlap between user U001 and Ad_A = |{technology, efficiency}| / |{technology, sports, efficiency, workplace}| = 2 / 4 = 0.5.

[0081] The overlap between user U001 and Ad_B = |{}| / |{technology, workplace, efficiency, life, leisure, white-collar}| = 0.

[0082] The overlap between user U001 and Ad_C = |{technology, workplace}| / |{technology, education, workplace, efficiency}| = 2 / 4 = 0.5.

[0083] The audience matching threshold was set based on historical campaign data analysis. The analysis showed that when the Jaccard similarity coefficient was below 0.25, the click-through rate and conversion rate of the ads dropped significantly, failing to cover costs. Therefore, the audience matching threshold was set to 0.25. Based on this threshold, the overlap between Ad_A and Ad_C (both 0.5) was higher than 0.25, while the overlap between Ad_B and Ad_C was 0, below the threshold. Therefore, Ad_A and Ad_C were selected, creating a candidate ad placement resource list [Ad_A, Ad_C].

[0084] The slot machine revenue assessment submodule calls the list of candidate resources and treats each resource item in the list as an independent arm of the multi-armed slot machine algorithm. Based on the historical conversion rate and deployment cost data of multiple resource items, the upper confidence limit algorithm is used to calculate the upper limit of the confidence interval of each arm to obtain the expected revenue value of the resource.

[0085] The slot machine revenue assessment submodule specifically includes:

[0086] For the i-th ad resource in the candidate ad resource list, collect its total number of historical ad placements after t rounds of placement. Total number of historical ad placements across all ad resources And calculate its historical average conversion rate. ;

[0087] Call historical average conversion rate Total number of historical deployments Total number of historical ad placements across all ad resources By using the preset upper confidence limit formula: Calculations are performed to obtain the expected revenue value of each advertising resource in the list of candidate advertising resources;

[0088] in, This represents the expected return on the i-th ad resource in the t-th round of deployment. This represents the index of the ad resource in the list of potential ad delivery resources. Represents the round of deployment;

[0089] The slot machine revenue assessment submodule calls this list of candidate resources and treats Ad_A and Ad_C as two independent arms in the multi-armed slot machine algorithm. Based on the historical conversion rates and deployment cost data of these two resource items, the submodule uses the upper confidence bound algorithm to calculate the upper limit of the confidence interval for each arm, i.e., the expected revenue value of the resource.

[0090] The expected return on this resource is determined using the upper confidence limit formula. The calculations are as follows. A detailed explanation of each parameter in the formula is provided below: This represents the expected revenue value of the i-th ad resource in the t-th round of deployment; i is the index of the ad resource in the list of candidate ad resources; t represents the current deployment round. This represents the historical average conversion rate of the i-th ad resource up to round t-1; This represents the total number of times the i-th ad resource has been deployed up to round t-1; This represents the total number of historical ad placements for all candidate ad resources up to round t-1. The calculation logic of this formula lies in balancing "utilization" and "exploration": the first term The first item is the utilization item, which tends to select ads with a good historical performance; the second item is the exploration item, whose value increases with the number of times an ad is selected. The increase in the number of times the ad is selected decreases, thus giving more opportunities to ads that are selected less often.

[0091] Before calculation, the system collects historical delivery data. Assume the current delivery is the [number]th [delivery]. Round-robin campaign, total number of historical ad placements for all ad resources. This is the sum of the past 200 rounds. Of these, ad Ad_A (index i=1) was served. Secondly, its historical average conversion rate (defined here as the percentage of purchases completed after a click) is: Ad Ad_C (index i=2) was served. Secondly, its historical average conversion rate is Total number of historical ad placements for all candidate ad resources. Substitute the parameter values ​​into the formula to calculate: For .for The advantage of the formula lies in the introduction of an exploration term. The system can dynamically identify and favor options with high potential despite limited historical data (such as Ad_C), avoiding prematurely abandoning potentially better ad resources based solely on current average conversion rates. The calculated expected return for Ad_A is 0.3928, and for Ad_C, it is 0.6198. This result indicates that, in the current round, considering both historical performance and exploration value, Ad_C is the option with higher expected returns. This result will directly serve as the basis for subsequent weight allocation.

[0092] The resource weight dynamic adjustment submodule sorts the resources in the list of selected advertising resources based on the expected revenue value of multiple advertising resources. The sorting result is used as the basis for weight allocation. The module recalculates the allocation ratio of multiple resources in the advertising budget through a normalization function, and integrates the resource identifier with the parameters of the new allocation ratio to generate a resource adjustment instruction.

[0093] The resource weight dynamic adjustment submodule sorts Ad_A and Ad_C based on the expected revenue values ​​of the resources mentioned above: Ad_C (0.6198) > Ad_A (0.3928). This sorting result serves as the basis for weight allocation. The submodule recalculates the allocation ratio of the two in the campaign budget using a normalization function. The normalization operation here is: taking the expected revenue value of each resource as its weight, and then dividing it by the sum of the expected revenue values ​​of all resources. The allocation ratio of Ad_C = (i.e., 61.21%). The allocation ratio of Ad_A = (i.e., 38.79%). Finally, the submodule integrates the resource identifier with the newly allocated ratio parameter to generate a resource adjustment instruction: {Instruction ID:CMD001, Ad Resource Allocation Weight:[{Ad_ID:Ad_C, Weight:0.6121},{Ad_ID:Ad_A, Weight:0.3879}], Targeting Time Adjustment: Evening Peak Hours (19:00-22:00), Targeting Priority Parameter: 1}. This instruction is then passed to the Interest Trend Analysis module.

[0094] Please see Figure 1 and Figure 5 The interest trend analysis module is used to obtain resource adjustment instructions, use interest trend clustering algorithm to classify node interest trends, and adjust subsequent advertising push parameters based on node category and resource adjustment instructions.

[0095] The node interest trend categories specifically refer to interest growth, interest decline, and interest fluctuation.

[0096] The interest trend analysis module includes:

[0097] The interest feature extraction submodule obtains resource adjustment instructions, collects user interaction logs of advertising resources in the instructions, counts the interaction frequency and duration data between each node and its associated advertisements, weights and fuses the statistical data, calculates the interest intensity of each node, and establishes a node interest feature vector.

[0098] The process of calculating the interest intensity of each node involves first normalizing the collected interaction frequency and duration data, and then using the formula... Perform calculations;

[0099] in, Represents the interest intensity of node n. This represents the normalized interaction frequency between node n and associated ads. The normalized interaction duration between node n and associated ads. Weighting coefficients representing normalized interaction frequencies. The weight coefficient represents the normalized interaction duration, where n is the node index, and satisfies... ;

[0100] The interest feature extraction submodule first obtains the instruction and, based on the ad resources Ad_C and Ad_A in the instruction, collects the interaction logs of users who interacted with these ads. For node U001, the submodule statistically analyzes its interaction data after receiving a new round of ad pushes. In the past 15 minutes, node U001 interacted with Ad_C (online programming course) 3 times (click to view details), with a total interaction duration of 90 seconds. To calculate the interest intensity of this node, the raw data is first normalized. Normalization uses the max-min normalization method, with the baseline max and min values ​​derived from statistical analysis of all users' interaction data with similar ads over the past hour. The statistics show that the maximum interaction frequency is 30 times / 15 minutes, and the minimum is 0; the maximum interaction duration is 600 seconds / 15 minutes, and the minimum is 0. The normalized interaction frequency of node U001 is shown below. Normalized interaction duration of node U001 The intensity of interest is calculated using the formula. Execution. The specific explanations of each parameter in the formula are as follows: The interest intensity represents node n (U001 in this case); It is the normalized interaction frequency of node n; It is the normalized interaction duration of node n; and These are the weighting coefficients for interaction frequency and interaction duration, respectively, and their sum is 1. The calculation logic of this formula is to obtain a more comprehensive and accurate measure of interest by weighting and fusing these two different dimensions of user engagement metrics: frequency and duration. Weighting coefficients and The weighting of AdC depends on the nature of the AdC. As an online course ad, the duration of user viewing time reflects their true learning intentions and potential conversion rate more accurately than simple click frequency. Therefore, interaction time should be given higher weight. To determine the specific value, an A / B test experiment was conducted, setting three different weightings.

[0101] ; ; The conversion rate of each group of users was tracked for a week.

[0102] Table 2. Weighting Coefficient Test Results

[0103]

[0104] As shown in Table 2, the weights are set to... and The experimental group achieved the highest final conversion rate of 6.8%. Therefore, this group's weights were selected by the system. The parameters were then substituted into the formula to calculate the interest intensity of node U001: The advantage of the formula lies in its weighting coefficients. and It is configurable and can be flexibly adjusted according to the goals of different advertising campaigns, so that the calculated interest intensity index is more aligned with business objectives. The result of 0.135 quantifies the level of interest of user U001 in ad Ad_C at the current time point. By repeating this calculation at consecutive time points, the system builds a node interest feature vector for each node (such as U001) consisting of continuous interest intensity values, for example [0.09, 0.11, 0.135].

[0105] The trend clustering analysis submodule calls the node interest feature vectors, applies the interest trend clustering algorithm, uses Euclidean distance as the similarity standard between vectors, it iteratively calculates and assigns the feature vectors to the nearest cluster center, and continuously updates the position of the cluster center until convergence, dividing the nodes in each cluster into a class, and obtaining the node interest trend category.

[0106] The trend clustering analysis submodule calls a batch of user node interest feature vectors and applies an interest trend clustering algorithm. The algorithm's execution process is as follows: First, three are randomly selected from all feature vectors as initial cluster centers, representing interest growth (e.g., [0.1, 0.3, 0.5]), interest decay (e.g., [0.5, 0.3, 0.1]), and interest fluctuation (e.g., [0.3, 0.5, 0.2]). Then, for each user's interest feature vector (e.g., [0.09, 0.11, 0.135] for U001), the algorithm uses Euclidean distance as a similarity metric to calculate its distance to these three cluster centers. The distance from the U001 vector to the growth cluster center is shown in the diagram. Distance from vector U001 to the decay class center Distance from vector U001 to the center of the wave class. Since vector U001 has the smallest distance to the center of the growth cluster, it is temporarily assigned to the interest growth cluster. After all vectors have been assigned, the algorithm recalculates the center of each cluster, taking the mean of all vectors within that cluster as the new center. This "assignment-update" process is iterated repeatedly until the change in the cluster center position is less than a minimum value (e.g., 1e-4) or the maximum number of iterations is reached, at which point the algorithm converges. Finally, the nodes within each cluster are uniformly classified into one category, resulting in the node interest trend category.

[0107] The push parameter adjustment submodule is based on the node interest trend category and resource adjustment instructions. For the interest trend represented by each category, it adjusts the direction and magnitude according to the weight in the instruction, quantitatively modifies the delivery frequency and target audience range in the ad push parameters, generates an instruction set of updated parameter values, and obtains subsequent ad push parameters.

[0108] The push parameter adjustment submodule quantifies and modifies subsequent ad push parameters based on the pre-defined node interest trend categories and previously received resource adjustment instructions. For users with increasing interest (including U001), these users exhibit a positive upward trend in interest. Instruction CMD001 has increased the resource weight of high-potential Ad_C to 61.21%. Based on this, the push parameter adjustment submodule further quantifies and modifies the ad delivery frequency for this user group, increasing the ad display frequency for Ad_C by 30% from the current baseline. The target audience remains unchanged. For users with declining interest, their interest intensity continues to decrease. The submodule reduces the delivery frequency for these users by 70% and slightly expands the definition of the target audience (e.g., adding a second-highest relevant tag to the user tags). For users with fluctuating interest, their interest performance is unstable. The submodule maintains the current delivery frequency and audience scope but switches the push ad creative library to another set (e.g., switching from "Feature Introduction" to "User Testimonials"). Through the above adjustments, the submodule generates an instruction set of updated parameter values, such as {User Group: Interest Growth Category, Ad ID: Ad_C, Frequency Control: +30%, Audience: Maintain}. This instruction set is the final parameter used for subsequent ad pushes.

[0109] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the scope of protection defined by the claims of the present invention.

Claims

1. An AI-driven short video ad placement optimization system, characterized in that, The system includes: The node feature extraction module is used to collect user behavior sequence parameters, content tags, and interest semantic parameters to obtain multi-dimensional node parameter data, input it into the BERT model for feature vectorization, call the node feature encoding results, and pass the node feature encoding results to the inflection point detection module. The inflection point detection module is used to obtain a node behavior feature sequence based on the node feature encoding result, call the node behavior feature sequence to input the temporal attention mechanism algorithm, detect and identify the node influence inflection point, and pass the node influence inflection point to the resource scheduling module. The inflection point detection module includes: The behavior feature serialization submodule extracts the timestamp information from each encoding result based on the node feature encoding result, reorders all node feature encodings according to the order of the timestamps, and concatenates the sorted encodings into a time-continuous data stream to generate a node behavior feature sequence. The temporal attention weight calculation submodule calls the node behavior feature sequence to input into the temporal attention mechanism algorithm, calculates the query vector and key vector for the feature vector of each node in the sequence, obtains the attention score through vector dot product operation, and then uses the Softmax function to normalize all scores to obtain the node temporal attention score. The influence turning point determination submodule performs a weighted summation of the node behavior feature sequence based on the node temporal attention score, calculates the context representation vector of each time node, monitors the change amplitude of the cosine similarity of the context representation vectors between adjacent nodes, compares the change amplitude with the set influence change threshold, identifies nodes whose change amplitude exceeds the threshold, and obtains the node influence turning point. The influence turning point determination submodule specifically includes: Collect the cosine similarity variation data of the context representation vectors between all adjacent nodes within a historical time window, and calculate the mean of the cosine similarity variation data. with standard deviation ; Call the mean The standard deviation and preset risk coefficient Through the dynamic threshold formula: The threshold for the change in influence is calculated. in, The threshold representing changes in influence. This is a preset risk coefficient constant; The resource scheduling module is used to obtain the node's influence inflection point, adjust the weight of the advertising resource parameter set using a multi-armed slot machine algorithm, and transmit the resource adjustment instruction to the interest trend analysis module. The resource scheduling module includes: The inflection point resource mapping submodule obtains the node influence inflection point, collects the set of advertising placement resource parameters, analyzes the user group profile associated with the inflection point and the target audience tags of multiple advertising resources, calculates the overlap between the two, and filters the advertising resources associated with the inflection point according to the preset audience matching threshold to establish a list of candidate placement resources. The slot machine revenue assessment submodule calls the list of candidate deployment resources, treats each resource item in the list as an independent arm of the multi-arm slot machine algorithm, and calculates the upper confidence interval of each arm based on the historical conversion rate and deployment cost data of multiple resource items, and obtains the expected revenue value of the resource. The resource weight dynamic adjustment submodule sorts the resources in the candidate placement resource list according to the expected revenue value of multiple advertising resources, uses the sorting result as the basis for weight allocation, recalculates the allocation ratio of multiple resources in the placement budget through a normalization function, and integrates the resource identifier with the parameters of the new allocation ratio to generate a resource adjustment instruction. The interest trend analysis module is used to obtain the resource adjustment instruction, use an interest trend clustering algorithm to classify the node interest trends, and adjust the subsequent advertising push parameters based on the node category and the resource adjustment instruction.

2. The AI-driven short video advertising optimization system according to claim 1, characterized in that, The node feature encoding results include user interest vectors, content relevance vectors, and behavioral preference vectors. The node influence inflection points specifically refer to node behavior change points, user decision threshold points, and influence fluctuation points. The resource adjustment instructions include advertising resource allocation weights, placement time adjustments, and placement priority parameters. The node interest trend categories specifically refer to interest growth, interest decay, and interest fluctuation.

3. The AI-driven short video advertising optimization system according to claim 2, characterized in that, The node feature extraction module includes: The behavior and interest data collection submodule collects user behavior sequence parameters, content tags, and interest semantic parameters. Based on user interaction logs, it calculates the number of clicks on multiple content tags and the average user dwell time. Combining the frequency of user feedback on differentiated interest semantics, it integrates multiple statistical data to generate raw user behavior indicators. The multi-dimensional node parameter construction submodule calls the original user behavior indicators, filters the user behavior sequence parameters based on the preset interaction activity benchmark value, removes sequences with interaction frequency lower than the benchmark value, and integrates the filtered behavior sequence parameters with content tags and interest semantic parameters in a structured manner to establish multi-dimensional node parameter data. The node feature vectorization submodule inputs the multi-dimensional node parameter data into the BERT model, uses the self-attention mechanism inside the model to calculate the correlation score between multiple dimension parameters, captures the deep contextual relationships in the parameter sequence through a multi-layer encoder structure, maps discrete node parameters to a continuous vector space, and obtains the node feature encoding results.

4. The AI-driven short video advertising optimization system according to claim 1, characterized in that, The interest trend analysis module includes: The interest feature extraction submodule obtains the resource adjustment instruction, collects the user interaction logs of the advertising resources in the instruction, counts the interaction frequency and duration data between each node and the associated advertisement, weights and fuses the statistical data, calculates the interest intensity of each node, and establishes a node interest feature vector. The trend clustering analysis submodule calls the node interest feature vector, applies the interest trend clustering algorithm, uses Euclidean distance as the measure of similarity between vectors, it iteratively calculates and assigns the feature vector to the nearest cluster center, and continuously updates the position of the cluster center until convergence, dividing the nodes in each cluster into a class, and obtaining the node interest trend category. The push parameter adjustment submodule, based on the node interest trend category and the resource adjustment instruction, adjusts the direction and magnitude of the interest trend represented by each category according to the weight in the instruction, quantitatively modifies the delivery frequency and target audience range in the ad push parameters, generates an instruction set of updated parameter values, and obtains subsequent ad push parameters.

5. The AI-driven short video advertising optimization system according to claim 4, characterized in that, The process of calculating the interest intensity of each node specifically involves first normalizing the collected interaction frequency and duration data, and then using the formula... Perform calculations; in, Represents the interest intensity of node n. This represents the normalized interaction frequency between node n and associated ads. The normalized interaction duration between node n and associated ads. Weighting coefficients representing normalized interaction frequencies. The weight coefficient represents the normalized interaction duration, where n is the node index, and satisfies... .

6. The AI-driven short video advertising optimization system according to claim 1, characterized in that, The slot machine revenue evaluation submodule specifically includes: For the i-th ad resource in the candidate ad resource list, collect its total historical ad placement count after t rounds of placement. Total number of historical ad placements across all ad resources And calculate its historical average conversion rate. ; Call the historical average conversion rate The total number of historical deployments Compared with the total number of historical ad deliveries for all the aforementioned ad resources Through the preset upper confidence limit formula Calculations are performed to obtain the expected revenue value of each advertising resource in the list of candidate advertising resources; in, This represents the expected return on the i-th ad resource in the t-th round of deployment. This represents the index of the ad resource in the list of potential ad delivery resources. This represents the round of deployment.

Images