An AI advertisement delivery decision system and method based on user behavior modeling

Abstract

This invention relates to the field of advertising placement decision technology, and discloses an AI advertising placement decision method based on user behavior modeling, comprising: collecting multimodal behavior data of users and constructing user behavior sequences; establishing a user behavior model based on the user behavior sequences; constructing an optimal advertising placement model based on the output results of the user behavior model and advertising placement requirements, and generating an optimal placement decision; displaying corresponding advertisements to users according to the optimal placement decision results, and recording feedback results for iteration.

Description

Technical Field

[0001] This invention relates to the field of advertising placement decision technology, specifically to an AI advertising placement decision system and method based on user behavior modeling. Background Technology

[0002] Digital advertising is an important part of the internet economy. Its core objective is to accurately deliver ads to the user groups most likely to convert (clicks, purchases, registrations, etc.) within a limited budget, thereby maximizing the return on investment for advertising.

[0003] Existing advertising delivery systems and methods still have many technical shortcomings, making it difficult to meet advertisers' multiple needs for delivery efficiency, conversion rates, and user experience. Specific deficiencies are as follows:

[0004] (1) Existing advertising delivery systems have a single data collection dimension, focusing only on one or a few user behaviors and ignoring implicit behaviors such as page dwell time and the collection and utilization of contextual information. This results in an incomplete portrayal of user interests and preferences, making it difficult to capture potential user needs and thus affecting the accuracy of delivery.

[0005] (2) Most user behavior modeling methods adopt a single-dimensional interest extraction approach, which fails to take into account both the immediacy of users' short-term interests and the stability of their long-term interests. Furthermore, the interest fusion process lacks a dynamic and controllable mechanism and cannot adaptively adjust the fusion weights according to changes in user behavior and contextual features, resulting in insufficient representation accuracy of the comprehensive interest vector.

[0006] (3) The existing ad placement decision generation process lacks a systematic multi-algorithm fusion mechanism, and relies on a single algorithm to screen candidate ads. It does not combine the association rules between user groups and ad categories, or the user's preference score for ads for comprehensive screening. Furthermore, it has not established a scientific comprehensive utility evaluation system, and cannot balance multiple objectives such as click-through rate, conversion rate, ad placement cost and brand enhancement. This results in insufficient optimization of ad placement decisions and makes it difficult to achieve efficient use of advertisers' budgets. Moreover, the existing ad placement system lacks a sound closed-loop iteration mechanism. The ad placement feedback data does not effectively feed back into the updates of user behavior models and ad placement decision models, resulting in the models being unable to adapt to the dynamic changes in user behavior. The ad placement strategy is difficult to continuously optimize, ultimately leading to problems such as high ad placement costs, low conversion efficiency, and strong user interference. It is impossible to achieve a win-win situation for advertisers, platforms and users.

[0007] Therefore, in view of the above-mentioned shortcomings of existing technologies, there is an urgent need for an AI advertising placement decision system and method that can comprehensively collect multimodal user behavior data and scientifically generate optimal placement decisions. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] To address the shortcomings of existing technologies, this application provides an AI-based advertising delivery decision-making system and method based on user behavior modeling.

[0010] (II) Technical Solution

[0011] To address the above problems, this application provides the following technical solution:

[0012] An AI-powered advertising delivery decision-making system based on user behavior modeling includes:

[0013] The data acquisition module is used to collect users' multimodal behavioral data in real time;

[0014] The user sequence construction module is used to construct user behavior sequences based on the data acquisition module.

[0015] The user behavior construction module is used to build a user behavior model based on the output of the user sequence construction module.

[0016] The ad delivery decision generation module is used to construct the optimal ad delivery model based on the output of the user behavior construction module and ad delivery requirements.

[0017] The delivery execution and feedback module is used to display corresponding advertisements to users based on the optimal delivery decision generated by the delivery decision generation module, record feedback results, and return the feedback results to the user sequence construction module for iterative updates.

[0018] An AI-based ad delivery decision-making method based on user behavior modeling includes:

[0019] Collect multimodal behavioral data of users and construct user behavior sequences;

[0020] Establish a user behavior model based on the user behavior sequence;

[0021] Based on the output of the user behavior model and the advertising needs, an optimal advertising delivery model is constructed to generate the optimal delivery decision.

[0022] Based on the optimal delivery decision, the corresponding advertisement is displayed to the user, and the feedback results are recorded for iteration.

[0023] Preferably, the step of collecting users' multimodal behavior data and constructing user behavior sequences specifically includes:

[0024] Real-time collection of multimodal behavior data of users on digital media platforms, including explicit behavior, implicit behavior and contextual information;

[0025] The multimodal behavioral data is preprocessed; the data preprocessing includes at least data cleaning, data standardization, and session segmentation.

[0026] User behavior sequences are constructed based on preprocessed multimodal behavior data. For each user, the behaviors within the session are arranged in ascending chronological order, and a sliding window is used to retain the K most recent behaviors, constructing a behavior sequence X=[x1, x2, ..., x...]. i , ..., x N ], where N and K are positive integers, and N≤K.

[0027] Preferably, the explicit behaviors include at least ad clicks, searches, browsing, favorites, purchases, adding to cart, and negative feedback behaviors; the negative feedback behaviors include at least blocking and uninterested click behaviors.

[0028] The implicit behaviors include at least page dwell time and scrolling depth;

[0029] The context information includes at least timestamps, device type, network environment, and geographical location.

[0030] Preferably, the step of establishing a user behavior model based on the user behavior sequence specifically includes:

[0031] x in the behavior sequence X i By mapping the embedding layer to a fixed-dimensional numerical vector, we obtain the embedding sequence E=[e1, e2, ..., e...]. i , ..., e N ];

[0032] Establish a dual-channel interest model; input the embedded sequence E into the Transformer model, and take the output of the last time step as the short-term interest vector S; maintain the user's long-term memory state L. t-1 The mean pooling of the embedded sequence E of the behavior is input into the GRU according to the session, and the long-term interest vector L is updated.

[0033] The short-term interest vector S and the long-term interest vector L are fused using a controllable gate vector to obtain the mathematical expression for the comprehensive interest vector I:

[0034] (1)

[0035] In formula (1), It is a controllable gate vector. For contextual features, b is a bilinear term, and b is a bias term. , and As weight;

[0036] Perform T random forward propagations on the data fusion process and calculate the empirical variance of the integrated interest vector. The mathematical expression is:

[0037] (2)

[0038] In formula (2), Let T be the mean of the comprehensive interest vectors described above;

[0039] The user behavior model outputs at least the comprehensive interest vector. Conversion rate (p) and user segmentation tags (y).

[0040] Preferably, the step of constructing an optimal advertising delivery model based on the user behavior model output and combining it with advertising delivery requirements to generate the optimal delivery decision specifically includes:

[0041] Obtain the output results of the user behavior model and the advertising delivery requirements; the advertising delivery requirements include at least the advertiser's budget, targeting requirements, and core objectives;

[0042] The FP-Growth algorithm is used to mine the association rules between user segmentation tags and ad categories offline, and the ad categories are filtered online based on the current user's user segmentation tags to generate the first candidate ad set A1;

[0043] A user-ad interaction matrix is ​​constructed using a collaborative filtering algorithm, and the target user's preference score for the ad is obtained. The collaborative filtering score is then weighted and corrected in conjunction with the output of the user behavior model to generate a second candidate ad set A2.

[0044] Take the union A = A1∪A2 of the first candidate ad set and the second candidate ad set. After filtering by advertiser targeting constraints, calculate the comprehensive utility score of each candidate ad. The mathematical expression for the comprehensive utility score of each candidate ad is:

[0045] (3)

[0046] In formula (3), For click-through rate, For conversion rate, To estimate costs, To improve the brand's score, , , and These are the weighting coefficients;

[0047] The ad with the highest overall utility score is selected as the optimal placement decision, and the target user, ad content, bid coefficient, frequency limit, and budget allocation are output.

[0048] Preferably, the step of offline mining of association rules between user segmentation tags and ad categories using the FP-Growth algorithm, and online filtering of matching ad categories based on the current user's user segmentation tags to generate a first candidate ad set A1, specifically includes:

[0049] In the offline phase, a transaction database is constructed using the user segmentation tags and advertising categories, and the FP-Growth algorithm is used to filter association rules that meet the minimum support ≥ 0.01 and minimum confidence ≥ 0.5.

[0050] During the online phase, the user segmentation tags of the current user are obtained, all matching rules are retrieved, and all advertisements under the corresponding advertisement category are selected as the first candidate advertisement set A1.

[0051] Preferably, the step of constructing a user-ad interaction matrix using a collaborative filtering algorithm, obtaining the target user's preference score for the ad, and then weighting and correcting the collaborative filtering score based on the output of the user behavior model to generate a second candidate ad set A2, specifically includes:

[0052] For the target users, construct a user-ad interaction matrix and obtain the target users' original scores on ad preferences;

[0053] The mathematical expression for correcting the original score by combining the comprehensive interest vector and conversion rate from the user behavior model is as follows:

[0054] (4)

[0055] In formula (4), The corrected preference score. The base preference scores predicted by the collaborative filtering algorithm. Embed vectors for advertising features;

[0056] The advertisement with the highest score is selected as candidate set A2.

[0057] Preferably, the step of displaying the corresponding advertisement to the user based on the optimal delivery decision result and recording the feedback results for iteration specifically includes:

[0058] By connecting to the advertising platform through the API interface, the optimal placement decision plan is encapsulated into the request format required by the advertising platform and sent to the corresponding platform to realize the fully automated execution of the advertising placement process;

[0059] Real-time collection of feedback data during the campaign process, including at least impressions, clicks, conversions, negative feedback, and actual budget expenditure;

[0060] Based on the collected feedback data, the user behavior model and the optimal advertising delivery model are iteratively updated.

[0061] (III) Beneficial Effects

[0062] Compared with existing technologies, this application provides an AI advertising delivery decision-making system and method based on user behavior modeling, which has the following beneficial effects:

[0063] 1. This method collects multimodal user behavior data in real time through a data acquisition module, covering explicit behavior, implicit behavior and contextual information. It breaks through the limitation of existing technologies that only collect single behavior data, and achieves comprehensive capture of user behavior, enabling more accurate mining of users' potential interests and needs.

[0064] 2. This method establishes a dual-channel interest model, using Transformer to capture short-term interests, GRU to maintain long-term memory, and controllable gate vectors for data fusion; it can effectively distinguish between users' immediate needs and long-term preferences, solving the problem that a single model cannot simultaneously take into account timeliness and stability.

[0065] 3. In the advertising decision-making stage, this method adopts a dual recall strategy of association rules and collaborative filtering. By taking the union of the results, it significantly expands the coverage of candidate ads. At the same time, through weighted correction and comprehensive utility score calculation, it ensures the accuracy and commercial value of the final placement decision.

[0066] 4. This method is designed with a complete feedback iteration, which transmits exposure, clicks, conversions and negative feedback data after the campaign to the model in real time; this enables the user behavior model and campaign strategy to be dynamically updated and optimized based on market feedback, realizing the system's self-evolution and effectively avoiding the problem of model performance decaying over time.

[0067] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0068] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0069] Figure 1 This is a schematic diagram of the AI ​​advertising placement decision-making method based on user behavior modeling proposed in this application;

[0070] Figure 2 This is a schematic diagram illustrating the principle of an AI advertising delivery decision system based on user behavior modeling, as described in this application. Detailed Implementation

[0071] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0072] The terms "first" and "second" in the specification and claims of this application may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise stated, "multiple" means two or more. Furthermore, "and / or" in the specification and claims indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0073] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicating the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0074] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0075] Please see Figures 1-2 This application provides a new technical solution: an AI advertising delivery decision-making system based on user behavior modeling, comprising:

[0076] The data acquisition module is used to collect users' multimodal behavioral data in real time;

[0077] The user sequence construction module is used to construct user behavior sequences based on the data acquisition module.

[0078] The user behavior construction module is used to build a user behavior model based on the output of the user sequence construction module.

[0079] The ad delivery decision generation module is used to construct the optimal ad delivery model based on the output of the user behavior construction module and ad delivery requirements.

[0080] The delivery execution and feedback module is used to display corresponding advertisements to users based on the optimal delivery decision generated by the delivery decision generation module, record feedback results, and return the feedback results to the user sequence construction module for iterative updates.

[0081] In a specific embodiment, the signal output terminal of the data acquisition module is connected to the signal input terminal of the user sequence construction module;

[0082] The signal output terminal of the user sequence construction module is connected to the signal input terminal of the user behavior construction module;

[0083] The signal output terminal of the user behavior construction module is connected to the signal output terminal of the delivery decision generation module;

[0084] The signal output terminal of the delivery decision generation module is connected to the signal input terminal of the delivery execution and feedback module, and the signal output terminal of the delivery execution and feedback module is connected to the signal input terminals of the user sequence construction module and the delivery decision generation module.

[0085] An AI-based ad delivery decision-making method based on user behavior modeling includes:

[0086] Collect multimodal behavioral data of users and construct user behavior sequences;

[0087] Establish a user behavior model based on the user behavior sequence;

[0088] Based on the output of the user behavior model and the advertising needs, an optimal advertising delivery model is constructed to generate the optimal delivery decision.

[0089] Based on the optimal delivery decision, the corresponding advertisement is displayed to the user, and the feedback results are recorded for iteration.

[0090] In this invention, the collection of users' multimodal behavior data and the construction of user behavior sequences specifically include:

[0091] Real-time collection of multimodal behavior data of users on digital media platforms, including explicit behavior, implicit behavior and contextual information;

[0092] The multimodal behavioral data is preprocessed; the data preprocessing includes at least data cleaning, data standardization, and session segmentation.

[0093] User behavior sequences are constructed based on preprocessed multimodal behavior data. For each user, the behaviors within the session are arranged in ascending chronological order, and a sliding window is used to retain the K most recent behaviors, constructing a behavior sequence X=[x1, x2, ..., x...]. i , ..., x N ], where N and K are positive integers, and N≤K.

[0094] In a specific embodiment, the data cleaning involves filtering duplicate behavior records, removing abnormal scrolling behaviors with a dwell time exceeding 10 minutes or less than 0.1 seconds, and removing invalid records with missing key fields.

[0095] The data standardization involves formatting each behavior record into a result-based tuple.

[0096] The session segmentation is based on the user dimension, with the dividing line being that the time interval between adjacent behaviors exceeds 30 minutes or the user actively closes the session / switches accounts, thus dividing the behavior into independent sessions;

[0097] In this invention, the explicit behaviors include at least ad clicks, searches, browsing, favorites, purchases, adding to cart, and negative feedback behaviors; the negative feedback behaviors include at least blocking and uninterested click behaviors.

[0098] The implicit behaviors include at least page dwell time and scrolling depth;

[0099] The context information includes at least timestamps, device type, network environment, and geographical location.

[0100] In this invention, the step of establishing a user behavior model based on the user behavior sequence specifically includes:

[0101] x in the behavior sequence X i By mapping the embedding layer to a fixed-dimensional numerical vector, we obtain the embedding sequence E=[e1, e2, ..., e...]. i , ..., e N ];

[0102] Establish a dual-channel interest model; input the embedded sequence E into the Transformer model, and take the output of the last time step as the short-term interest vector S; maintain the user's long-term memory state L. t-1 The mean pooling of the embedded sequence E of the behavior is input into the GRU according to the session, and the long-term interest vector L is updated.

[0103] The short-term interest vector S and the long-term interest vector L are fused using a controllable gate vector to obtain the mathematical expression for the comprehensive interest vector I:

[0104] (1)

[0105] In formula (1), It is a controllable gate vector. For contextual features, b is a bilinear term, and b is a bias term. , and As weight;

[0106] Perform T random forward propagations on the data fusion process and calculate the empirical variance of the integrated interest vector. The mathematical expression is:

[0107] (2)

[0108] In formula (2), Let T be the mean of the comprehensive interest vectors described above;

[0109] The user behavior model outputs at least the comprehensive interest vector. Conversion rate (p) and user segmentation tags (y).

[0110] In a specific embodiment, Monte Carlo Dropout is used to perform T random forward propagations on the interest fusion process during the inference phase, where T ≥ 30. The Dropout layer is kept active in each forward propagation, and the empirical variance of the combined interest vector output by the T propagations is calculated. .

[0111] In a specific embodiment, the conversion probability The MLP is a two-layer fully connected network with sigmoid activation in the output layer; the user group label y is obtained by clustering the interest preference vectors of all users using the offline K-Means clustering algorithm.

[0112] In this invention, the step of constructing an optimal advertising delivery model based on the user behavior model output and combining it with advertising delivery requirements to generate an optimal delivery decision specifically includes:

[0113] Obtain the output results of the user behavior model and the advertising delivery requirements; the advertising delivery requirements include at least the advertiser's budget, targeting requirements, and core objectives;

[0114] The FP-Growth algorithm is used to mine the association rules between user segmentation tags and ad categories offline, and the ad categories are filtered online based on the current user's user segmentation tags to generate the first candidate ad set A1;

[0115] A user-ad interaction matrix is ​​constructed using a collaborative filtering algorithm, and the target user's preference score for the ad is obtained. The collaborative filtering score is then weighted and corrected in conjunction with the output of the user behavior model to generate a second candidate ad set A2.

[0116] Take the union A = A1∪A2 of the first candidate ad set and the second candidate ad set. After filtering by advertiser targeting constraints, calculate the comprehensive utility score of each candidate ad. The mathematical expression for the comprehensive utility score of each candidate ad is:

[0117] (3)

[0118] In formula (3), For click-through rate, For conversion rate, To estimate costs, To improve the brand's score, , , and These are the weighting coefficients;

[0119] The ad with the highest overall utility score is selected as the optimal placement decision, and the target user, ad content, bid coefficient, frequency limit, and budget allocation are output.

[0120] In a specific embodiment, the advertiser's budget includes at least a total budget and a daily budget cap; the advertiser's core objectives include at least maximum conversion rate and increased brand exposure; and the advertiser's targeting requirements include at least geographic region, age group, and device type.

[0121] In this invention, the offline mining of association rules between user segmentation tags and advertising categories using the FP-Growth algorithm, and the online filtering of matching advertising categories based on the current user's user segmentation tags to generate a first candidate advertising set A1, specifically includes:

[0122] In the offline phase, a transaction database is constructed using the user segmentation tags and advertising categories, and the FP-Growth algorithm is used to filter association rules that meet the minimum support ≥ 0.01 and minimum confidence ≥ 0.5.

[0123] During the online phase, the user segmentation tags of the current user are obtained, all matching rules are retrieved, and all advertisements under the corresponding advertisement category are selected as the first candidate advertisement set A1.

[0124] In this invention, the method of constructing a user-ad interaction matrix using a collaborative filtering algorithm, obtaining the target user's preference score for the ad, and then weighting and correcting the collaborative filtering score based on the output of the user behavior model to generate a second candidate ad set A2, specifically includes:

[0125] For the target users, construct a user-ad interaction matrix and obtain the target users' original scores on ad preferences;

[0126] The mathematical expression for correcting the original score by combining the comprehensive interest vector and conversion rate from the user behavior model is as follows:

[0127] (4)

[0128] In formula (4), The corrected preference score. The base preference scores predicted by the collaborative filtering algorithm. Embed vectors for advertising features;

[0129] The advertisement with the highest score is selected as candidate set A2.

[0130] In a specific embodiment, a user-advertising interaction matrix is ​​constructed. M represents the total number of users, and N represents the total number of advertisements; each element in the matrix... This represents the historical interaction strength of user u with advertisement a.

[0131] In this invention, the step of displaying the corresponding advertisement to the user based on the optimal delivery decision result and recording the feedback results for iteration specifically includes:

[0132] By connecting to the advertising platform through the API interface, the optimal placement decision plan is encapsulated into the request format required by the advertising platform and sent to the corresponding platform to realize the fully automated execution of the advertising placement process;

[0133] Real-time collection of feedback data during the campaign process, including at least impressions, clicks, conversions, negative feedback, and actual budget expenditure;

[0134] Based on the collected feedback data, the user behavior model and the optimal advertising delivery model are iteratively updated.

[0135] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0136] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based advertising delivery decision-making system based on user behavior modeling, characterized in that, include: The data acquisition module is used to collect users' multimodal behavioral data in real time; The user sequence construction module is used to construct user behavior sequences based on the data acquisition module. The user behavior construction module is used to build a user behavior model based on the output of the user sequence construction module. The ad delivery decision generation module is used to construct the optimal ad delivery model based on the output of the user behavior construction module and ad delivery requirements. The delivery execution and feedback module is used to display corresponding advertisements to users based on the optimal delivery decision generated by the delivery decision generation module, record feedback results, and return the feedback results to the user sequence construction module for iterative updates.

2. An AI-based advertising delivery decision-making method based on user behavior modeling, characterized in that, include: Collect multimodal behavioral data of users and construct user behavior sequences; Establish a user behavior model based on the user behavior sequence; Based on the output of the user behavior model and the advertising needs, an optimal advertising delivery model is constructed to generate the optimal delivery decision. Based on the optimal delivery decision, the corresponding advertisement is displayed to the user, and the feedback results are recorded for iteration.

3. The AI ​​advertising placement decision-making method based on user behavior modeling according to claim 2, characterized in that, The process of collecting multimodal user behavior data and constructing user behavior sequences specifically includes: Real-time collection of multimodal behavior data of users on digital media platforms, including explicit behavior, implicit behavior and contextual information; The multimodal behavioral data is preprocessed; the data preprocessing includes at least data cleaning, data standardization, and session segmentation. Based on the pre-processed multi-modal behavior data, a user behavior sequence is constructed; for each user, the behaviors within the session are arranged in chronological order, and a sliding window is used to retain the latest K behaviors, to construct a behavior sequence X=[x1, x2, …, x i , …, x N ], where N and K are positive integers, and N≤K.

4. The AI ​​advertising placement decision-making method based on user behavior modeling according to claim 3, characterized in that, The explicit behaviors include at least ad clicks, searches, browsing, favorites, purchases, adding to cart, and negative feedback behaviors; the negative feedback behaviors include at least blocking and "not interested" click behaviors. The implicit behaviors include at least page dwell time and scrolling depth; The context information includes at least timestamps, device type, network environment, and geographical location.

5. The AI ​​advertising placement decision-making method based on user behavior modeling according to claim 3, characterized in that, The step of establishing a user behavior model based on the user behavior sequence specifically includes: x in the behavior sequence X i By mapping the embedding layer to a fixed-dimensional numerical vector, we obtain the embedding sequence E=[e1, e2, ..., e...]. i , ..., e N ]; Establish a dual-channel interest model; input the embedded sequence E into the Transformer model, and take the output of the last time step as the short-term interest vector S; maintain the user's long-term memory state L. t-1 The mean pooling of the embedded sequence E of the behavior is input into the GRU according to the session, and the long-term interest vector L is updated. The short-term interest vector S and the long-term interest vector L are fused using a controllable gate vector to obtain the mathematical expression for the comprehensive interest vector I: （1） In formula (1), It is a controllable gate vector. For contextual features, b is a bilinear term, and b is a bias term. , and As weight; Perform T random forward propagations on the data fusion process and calculate the empirical variance of the integrated interest vector. The mathematical expression is: （2） In formula (2), Let T be the mean of the comprehensive interest vectors described above; The user behavior model outputs at least the comprehensive interest vector. Conversion rate (p) and user segmentation tags (y).

6. The AI ​​advertising placement decision-making method based on user behavior modeling according to claim 5, characterized in that, The step of constructing an optimal advertising delivery model based on the user behavior model output and combining it with advertising delivery requirements to generate the optimal delivery decision specifically includes: Obtain the output results of the user behavior model and the advertising delivery requirements; the advertising delivery requirements include at least the advertiser's budget, targeting requirements, and core objectives; The FP-Growth algorithm is used to mine the association rules between user segmentation tags and ad categories offline, and the ad categories are filtered online based on the current user's user segmentation tags to generate the first candidate ad set A1; A user-ad interaction matrix is ​​constructed using a collaborative filtering algorithm, and the target user's preference score for the ad is obtained. The collaborative filtering score is then weighted and corrected in conjunction with the output of the user behavior model to generate a second candidate ad set A2. Take the union A = A1∪A2 of the first candidate ad set and the second candidate ad set. After filtering by advertiser targeting constraints, calculate the comprehensive utility score of each candidate ad. The mathematical expression for the comprehensive utility score of each candidate ad is: （3） In formula (3), For click-through rate, For conversion rate, To estimate costs, To improve the brand's score, , , and These are the weighting coefficients; The ad with the highest overall utility score is selected as the optimal placement decision, and the target user, ad content, bid coefficient, frequency limit, and budget allocation are output.

7. The AI ​​advertising placement decision-making method based on user behavior modeling according to claim 6, characterized in that, The process involves offline mining of association rules between user segmentation tags and ad categories using the FP-Growth algorithm, and online filtering of matching ad categories based on the current user's user segmentation tags to generate a first candidate ad set A1. Specifically, this includes: In the offline phase, a transaction database is constructed using the user segmentation tags and advertising categories, and the FP-Growth algorithm is used to filter association rules that meet the minimum support ≥ 0.01 and minimum confidence ≥ 0.

5. During the online phase, the user segmentation tags of the current user are obtained, all matching rules are retrieved, and all advertisements under the corresponding advertisement category are selected as the first candidate advertisement set A1.

8. The AI ​​advertising placement decision-making method based on user behavior modeling according to claim 6, characterized in that, The process involves constructing a user-ad interaction matrix using a collaborative filtering algorithm, obtaining the target user's preference score for the ad, and then weighting and correcting the collaborative filtering score based on the output of the user behavior model to generate a second candidate ad set A2. Specifically, this includes: For the target users, construct a user-ad interaction matrix and obtain the target users' original scores on ad preferences; The mathematical expression for correcting the original score by combining the comprehensive interest vector and conversion rate from the user behavior model is as follows: （4） In formula (4), The corrected preference score. The base preference scores predicted by the collaborative filtering algorithm. Embed vectors for advertising features; The advertisement with the highest score is selected as candidate set A2.

9. The AI ​​advertising placement decision-making method based on user behavior modeling according to claim 2, characterized in that, The step of displaying corresponding ads to users based on the optimal placement decision and recording feedback results for iteration specifically includes: By connecting to the advertising platform through the API interface, the optimal placement decision plan is encapsulated into the request format required by the advertising platform and sent to the corresponding platform to realize the fully automated execution of the advertising placement process; Real-time collection of feedback data during the campaign process, including at least impressions, clicks, conversions, negative feedback, and actual budget expenditure; Based on the collected feedback data, the user behavior model and the optimal advertising delivery model are iteratively updated.

Images