CN-121094896-B - AI large model driven short video advertisement putting optimizing system

Abstract

The invention relates to the technical field of machine learning, in particular to an AI large model driven short video advertisement delivery optimization system, which comprises a node characteristic extraction module, a turning point detection module, a resource scheduling module and an interest trend analysis module. According to the invention, through deep semantic vectorization of multi-dimensional data, dynamic capture of user behaviors and content interests is combined, key turning of node influence is perceived, fine-grained classification is carried out on the user interest change trend, advertisement pushing parameters are dynamically adjusted, real-time weight optimization of a resource allocation strategy is realized, high adaptation of advertisement pushing content and user interests is realized, continuous promotion of advertisement putting effect is promoted, advertisement resource utilization rate is improved, user experience is positively promoted, cooperation from multi-dimensional data acquisition to putting strategy optimization is realized in the whole process, user interest deviation is responded quickly, advertisement matching and interactive performance are enhanced, and benefit promotion and data value release are brought to short video advertisement service.

Inventors

• PENG WENPING

Assignees

• 今日智联(武汉)信息技术有限公司

Dates

Publication Date

20260512

Application Date

20250801

Claims (6)

1. An ai large model driven short video advertisement delivery optimization system, comprising: The node characteristic extraction module is used for acquiring user behavior sequence parameters, content labels and interest semantic parameters, acquiring multidimensional node parameter data, inputting the multidimensional node parameter data into the BERT model for characteristic vectorization, calling a node characteristic coding result, and transmitting the node characteristic coding result to the turning point detection module; The turning point detection module is used for acquiring a node behavior feature sequence based on the node feature coding result, calling the node behavior feature sequence to be input into a time sequence attention mechanism algorithm, detecting and judging a node influence turning point, and transmitting the node influence turning point to the resource scheduling module; the turning point detection module comprises: The behavior feature serialization submodule extracts time stamp information in each coding result based on the node feature coding result, reorders all node feature codes according to the sequence of the time stamps, and connects the ordered codes in series to form a time-continuous data stream to generate a node behavior feature sequence; the time sequence attention weight calculation submodule calls the node behavior feature sequence to be input into a time sequence attention mechanism algorithm, calculates a query vector and a key vector for the feature vector of each node in the sequence, obtains attention scores through vector dot product operation, and then normalizes all scores by using a Softmax function to obtain the time sequence attention score of the node; the influence turning judging submodule carries out weighted summation on the node behavior feature sequences according to the node time sequence attention score, calculates a context representation vector of each time node, monitors cosine similarity variation amplitude of the context representation vector between adjacent nodes, compares the variation amplitude with a set influence variation threshold, and identifies nodes with variation amplitude exceeding the threshold to obtain node influence turning points; The influence turning judging submodule specifically comprises: collecting cosine similarity variation amplitude data of context expression vectors among all adjacent nodes in a historical time window, and calculating the mean value of the cosine similarity variation amplitude data And standard deviation of ; Invoking the mean value Said standard deviation Preset risk factor By the dynamic threshold formula: calculating to obtain the influence change threshold; Wherein, the Representing the threshold value of the change in the influence, A preset risk coefficient constant; The resource scheduling module is used for acquiring the node influence turning points, carrying out weight adjustment on the advertisement putting resource parameter set by adopting a multi-arm slot machine algorithm, and transmitting a resource adjustment instruction to the interest trend analysis module; The resource scheduling module comprises: The turning point resource mapping sub-module acquires the turning points of the node influence, acquires the advertisement delivery resource parameter set, analyzes the user group portraits associated with the turning points and the target audience labels of a plurality of advertisement resources, calculates the coincidence ratio of the user group portraits and the target audience labels, screens the advertisement resources associated with the turning points according to a preset audience matching degree threshold value, and establishes a to-be-selected delivery resource list; The slot machine income evaluation sub-module calls the to-be-selected put resource list, takes each resource item in the list as an independent arm of the multi-arm slot machine algorithm, calculates the upper limit of a confidence interval of each arm by using an upper confidence limit algorithm according to the historical conversion rate and the put cost data of a plurality of resource items, and obtains a resource expected income value; the resource weight dynamic adjustment submodule sorts the resources in the to-be-selected delivery resource list according to the expected resource income values of the plurality of advertisement resources, takes the sorting result as the basis of weight distribution, recalculates the distribution proportion of the plurality of resources in the delivery budget through a normalization function, integrates the resource identification with the parameters of the new distribution proportion, and generates a resource adjustment instruction; And the interest trend analysis module is used for acquiring the resource adjustment instruction, classifying the node interest trend by adopting an interest trend clustering algorithm, and adjusting the follow-up advertisement pushing parameters based on the node category and the resource adjustment instruction.
2. 2. The AI large model driven short video advertisement delivery optimization system of claim 1, wherein the node feature encoding results comprise a user interest vector, a content correlation vector, a behavior preference vector, the node influence turning points are specifically node behavior change points, user decision threshold points, influence fluctuation points, the resource adjustment instructions comprise advertisement resource allocation weights, delivery period adjustment, delivery priority parameters, and the node interest trend category is specifically an interest increase category, an interest decay category, and an interest fluctuation category.
3. 3. The AI large model driven short video advertisement delivery optimization system of claim 2, wherein the node feature extraction module comprises: The behavior and interest data acquisition submodule acquires user behavior sequence parameters, content tags and interest semantic parameters, calculates clicking times of a plurality of content tags and average residence time of a user based on a user interaction log, integrates a plurality of items of statistical data by combining feedback frequency of the user on differentiated interest semantics, and generates a user original behavior index; the multidimensional node parameter construction submodule invokes the original behavior index of the user, screens the user behavior sequence parameters according to a preset interaction activity reference value, eliminates sequences with interaction frequency lower than the reference value, and carries out structural integration on the screened behavior sequence parameters, content tags and interest semantic parameters to establish multidimensional node parameter data; The node characteristic vectorization sub-module inputs the multidimensional node parameter data into a BERT model, calculates relevance scores among a plurality of dimensional parameters by using a self-attention mechanism in the model, captures deep context relations in a parameter sequence through a multi-layer encoder structure, maps discrete node parameters into a continuous vector space and acquires a node characteristic coding result.
4. 4. The AI large model driven short video advertising optimization system of claim 1, wherein the interest trend analysis module comprises: The interest feature extraction submodule acquires the resource adjustment instruction, acquires a user interaction log of advertisement resources in the instruction, counts the interaction frequency and duration data of the advertisement resources and associated advertisements for each node, carries out weighted fusion on the statistical data, calculates the interest intensity of each node, and establishes a node interest feature vector; The trend cluster analysis submodule calls the node interest feature vectors, applies an interest trend clustering algorithm, takes Euclidean distance as a measurement standard of similarity among vectors, distributes the feature vectors to the nearest cluster center through iterative calculation, continuously updates the positions of the cluster centers until convergence, and divides the nodes in each cluster into a class to obtain node interest trend categories; The pushing parameter adjustment submodule carries out quantitative modification on the frequency of putting advertisements in the advertisement pushing parameters and the range of target people according to the weight adjustment direction and the amplitude in the instruction aiming at the interest trend represented by each category based on the node interest trend category and the resource adjustment instruction, generates an instruction set of updated parameter values, and acquires the follow-up advertisement pushing parameters.
5. 5. The AI large model driven short video advertisement delivery optimization system as set forth in claim 4, wherein the calculation of interest intensity of each node comprises normalizing the collected interaction frequency and time length data, and then passing through the formula Calculating; Wherein, the Representing the strength of interest of the node n, Normalized interaction frequency of representative node n with the associated advertisement, A normalized interaction time period representing node n with the associated advertisement, The weight coefficients representing the normalized interaction frequency, Weight coefficient representing normalized interaction time length, n is node index and satisfies 。
6. 6. The AI large model driven short video advertising optimization system of claim 1, wherein the slot machine revenue evaluation submodule specifically comprises: aiming at the ith advertisement resource in the to-be-selected put resource list, collecting the historical total number of times of putting after t rounds of putting Together with the historical total number of impressions of all advertising resources And calculate the historical average conversion ; Invoking the historical average conversion rate The total number of times of the historical delivery Together with the historical total times of putting all the advertising resources Through a preset upper confidence limit formula Calculating to obtain a resource expected benefit value of each advertisement resource in the to-be-selected delivery resource list; Wherein, the Representing the expected revenue value of the ith advertising resource at the time of the t-th impression, Representing the advertising resource index in the list of impression resources to be selected, Representing the run-in round.

Description

AI large model driven short video advertisement putting optimizing system Technical Field The invention relates to the technical field of machine learning, in particular to a short video advertisement delivery optimization system driven by an AI large model. Background The technical field of machine learning relates to automatic modeling and prediction of complex problems by using a data-driven method, and the field comprises core matters such as feature engineering model training model evaluation and optimization, is widely applied to a plurality of actual scenes such as advertisement putting of an image recognition voice recognition natural language processing recommendation system, is based on large-scale data, and is used for mining potential rules in the data through a training model and being applied to actual business decision scenes. The conventional short video advertisement delivery optimization system is used for optimizing the advertisement delivery on a short video platform by adopting modes of click rate prediction conversion rate prediction user portrait analysis, advertisement sequencing and the like based on multidimensional data such as user behavior historical interest preference content characteristic advertisement materials, and the common practice comprises scoring the matching degree of users and advertisements by using a machine learning model and sequencing according to the score, and predicting the performances of different advertisement materials or delivery strategies in a target user group through a historical advertisement click behavior or conversion behavior data training model, so that reasonable distribution of advertisement delivery resources and dynamic adjustment of delivery contents are realized. In the prior art, model training and matching are often carried out by depending on historical behaviors and interest preference, due to lack of sensitive perception of user interest change nodes, certain hysteresis exists in resource allocation, advertisement delivery is difficult to respond to focus of attention of a user in time under the scene of rapid interest fluctuation or content hot spot fluctuation, advertisement pushing homogenization is caused, personalized recommendation effect is affected, a part of advertisement resources are distributed inefficiently, user experience is reduced due to mismatch of advertisements and actual interests, when a short video platform has a hot topic, interest turning is not captured timely by a system, advertisement pushing content is disjointed with real-time requirements of the user, and advertisement delivery value and platform income are affected. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a short video advertisement delivery optimization system driven by an AI large model. In order to achieve the above purpose, the invention adopts the following technical scheme that the AI large model driven short video advertisement putting optimization system comprises: The node characteristic extraction module is used for acquiring user behavior sequence parameters, content labels and interest semantic parameters, acquiring multidimensional node parameter data, inputting the multidimensional node parameter data into the BERT model for characteristic vectorization, calling a node characteristic coding result, and transmitting the node characteristic coding result to the turning point detection module; The turning point detection module is used for acquiring a node behavior feature sequence based on the node feature coding result, calling the node behavior feature sequence to be input into a time sequence attention mechanism algorithm, detecting and judging a node influence turning point, and transmitting the node influence turning point to the resource scheduling module; The resource scheduling module is used for acquiring the node influence turning points, carrying out weight adjustment on the advertisement putting resource parameter set by adopting a multi-arm slot machine algorithm, and transmitting the resource adjustment instruction to the interest trend analysis module; And the interest trend analysis module is used for acquiring the resource adjustment instruction, classifying the node interest trend by adopting an interest trend clustering algorithm, and adjusting the follow-up advertisement pushing parameters based on the node category and the resource adjustment instruction. As a further scheme of the invention, the node characteristic coding result comprises a user interest vector, a content correlation vector and a behavior preference vector, wherein the node influence turning point is specifically a node behavior change point, a user decision threshold point and an influence fluctuation point, the resource adjustment instruction comprises advertisement resource allocation weight, delivery period adjustment and delivery priority parameters, and the node interest trend category is s