Advertisement design assisting method and system based on artificial intelligence

Abstract

The invention discloses an advertisement design assistance method and system based on artificial intelligence, and relates to the technical field of intelligent design assistance, and the method comprises the steps: carrying out the data collection of a target user for advertisement design, constructing a behavior matrix, determining an earliest time point as a reference point, carrying out the standard deviation analysis, and determining a diffusion parameter; time spread values of user interests are analyzed, and a single time spread vector is calculated. According to the method, a local activation factor based on advertisement composite embedding is introduced, so that the weight of each behavior event can be dynamically associated with the difference of advertisement characteristics, user behaviors are divided according to users to which the user behaviors belong, and interest distribution is counted in combination with a distribution matrix subjected to space-time weighting, so that the user behavior analysis accuracy is improved. And further reducing into a user specific interest vector, generating a category number-based distribution vector for the user interest graph by combining a width-depth model, and further defining an aggregation proportion of different interest categories to user behaviors and advertisement contents.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent design assistance, in particular to an advertisement design assistance method and system based on artificial intelligence. BACKGROUND

[0002] The rapid growth of artificial intelligence, especially deep learning technology, provides a new solution for advertisement design, which enables the generation of advertisement content to be intelligently optimized in combination with multi-modal data (such as text, images and category information) and user behavior. Typically, pre-trained models are used to extract semantic features of advertisement text or image recognition technology is used to extract visual features of advertisements, which are processed in a single dimension to achieve a preliminary adaptation of advertisements to user preferences. However, traditional advertisement design assistance technology still lacks a deep understanding and systematic support for the temporal dynamic changes of user behavior, the content relevance between behavior events and the hierarchical optimization of advertisements for specific interest groups. The temporal characteristics of user behavior are usually processed using fixed time windows or simple periodic segmentation methods in existing systems, which ignores the dynamic evolution process of behavior events over time and reduces the accuracy of the trend of changes in user interest. SUMMARY

[0003] In view of the above existing problems, the present application is proposed.

[0004] Therefore, the present application provides an advertisement design assistance method and system based on artificial intelligence to solve the problem that the temporal characteristics of user behavior are usually processed using fixed time windows or simple periodic segmentation methods in existing systems, which ignores the dynamic evolution process of behavior events over time and reduces the accuracy of the trend of changes in user interest.

[0005] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides an advertisement design assistance method based on artificial intelligence, comprising: Data collection for target users of advertisement design, construction of a behavior matrix, determination of the earliest time point as a reference point, standard deviation analysis to determine the diffusion parameter, analysis of the time diffusion value of user interest, calculation of a single time diffusion vector, measurement of the content relevance between behavior events using cosine similarity, construction of a spatio-temporal distribution matrix combining the time diffusion value, extraction of advertisement features to form a composite feature embedding, dynamic calculation of the weight of behavior events combining the behavior matrix, weighting of the spatio-temporal distribution matrix to obtain a weighted distribution matrix, and weighted calculation of the user interest vector for a single user's behavior. Using the user interest vector and the weighted distribution matrix, the reference behavior time interval and the content similarity, a cross feature matrix is generated and converted into a wide model feature fusion, the behavior embedding result is combined with the advertisement feature in a nonlinear manner as an embedding vector, the output interest matching distribution is predicted, the optimization result of the wide deep model is calculated through a nonlinear activation, and the user interest distribution is used as an output; By comparing the user interest distribution with the advertisement feature one by one, a user interest response matrix is generated, the contribution degree of each advertisement feature is analyzed, the response degree of the user interest distribution on different advertisement features is quantified, and the overall performance of the advertisement is quantified through a comprehensive index, and the distribution of the advertisement performance on different users is further quantified. The target user is analyzed in terms of interest, and all the features of the advertisement are prioritized to form a contribution feature table.

[0006] As a preferred scheme of the advertisement design auxiliary method based on artificial intelligence, wherein: the time diffusion value of the user interest is analyzed, a single time diffusion vector is calculated, the content correlation between behavior events is measured by using cosine similarity, and the time diffusion value is combined to construct a space-time distribution matrix, a composite feature embedding is extracted from the advertisement features, the weight of the behavior event is dynamically calculated by combining the behavior matrix, and the weighted distribution matrix is obtained by weighting the space-time distribution matrix, and the user interest vector is calculated by weighting the behavior of a single user, including, Based on the target user of the advertisement design, user behavior data is collected to construct a behavior matrix; The earliest time point of all user behaviors in the behavior matrix is selected as a reference point, and the standard deviation analysis is performed on all user behavior time intervals, and then the diffusion parameters suitable for the diffusion range are determined; The time diffusion process of the user interest is modeled as a quasi-ballistic diffusion, the dynamic diffusion range of each behavior event in the behavior matrix is calculated, and the time diffusion value is combined into a single time diffusion vector; The content correlation between behavior events is measured by using cosine similarity, and the time diffusion value is combined to calculate the space-time distribution value and construct a space-time distribution matrix; The pre-trained word embedding model is used to extract the advertisement text data of the advertisement content, the deep learning image coding is used to generate the image feature vector of the advertisement image, the actual category data of the advertisement is collected, and the One-Hot coding and embedding are performed; The composite feature embedding of the target advertisement is constructed by integrating the advertisement text data, the image feature vector and the category data, the weight of the behavior event is dynamically calculated as a local activation factor by combining the behavior matrix, and the space-time distribution matrix is weighted; Different behaviors of different users in the behavior matrix are obtained, and the behaviors of a single user are counted, and the user behavior interest vector is calculated by weighting.

[0007] As a preferred scheme of the advertisement design auxiliary method based on artificial intelligence, wherein: the user interest vector and the weighted distribution matrix are used to generate a cross feature matrix by referring to the behavior time interval and the content similarity, and the behavior embedding result and the advertisement features are nonlinearly combined as an embedding vector, the output interest matching distribution is predicted, the optimization result of the wide-deep model is calculated through nonlinear activation, and the user interest distribution is obtained, including, The user interest vector and the weighted distribution matrix are used to generate a cross feature matrix by referring to the behavior time interval and the content similarity, and the behavior content and the time are processed in combination, the cross feature matrix is converted into a feature fusion of a wide model through weighted summation; A standard sequence embedding framework based on user behavior sequences is adopted, a Transformer is used to model the time sequence, and the time diffusion characteristics are integrated into the sequence embedding, the behavior embedding result and the advertisement features are nonlinearly combined as an embedding vector; A deep model is constructed by using a fully connected feedforward neural network, and is pre-trained according to historical data, an input vector is composed of the embedding vector, the composite feature embedding and the user behavior interest vector after normalization, and the output interest matching distribution is predicted; The interest matching distribution is combined with the linear feature output of the wide model, the optimization result of the wide-deep model is calculated through nonlinear activation, and the user interest distribution is obtained.

[0008] As a preferred scheme of the advertisement design auxiliary method based on artificial intelligence, wherein: the user interest response matrix is generated, the contribution degree of each advertisement feature is analyzed, the response degree of the user interest distribution on different advertisement features is quantified, the overall effectiveness of the advertisement is quantified through a comprehensive index, and the distribution of the advertisement effectiveness on different users is further quantified, including, The user interest response matrix is generated by comparing the user interest distribution with the advertisement features one by one, the advertisement features of the user interest response matrix are normalized, and the contribution degree of each advertisement feature is analyzed; The contribution value of the advertisement feature is used to quantify the response degree of the user interest distribution on different advertisement features, the overall effectiveness of the advertisement is quantified through a comprehensive index, and the distribution of the advertisement effectiveness on different users is further quantified in addition to the overall effectiveness of a single advertisement.

[0009] As a preferred scheme of the advertisement design auxiliary method based on artificial intelligence, wherein: the interest layer analysis of the target user includes, According to the interest response effectiveness data of the user to the advertisement, the interest threshold value is defined based on historical data, the mean and two standard deviations of the historical interest effectiveness are determined, the user whose interest response effectiveness is greater than or equal to the interest threshold value is marked as a high-interest user, the user whose interest response effectiveness is less than the interest threshold value is marked as a low-interest user, and the target user is subjected to interest hierarchical analysis.

[0010] As a preferred scheme of the advertisement design auxiliary method based on artificial intelligence, wherein: the priority of all features of the advertisement is sorted to form a contribution feature table, including, Based on the contribution degree of the advertisement features, the priority of all features of the advertisement is sorted, the contribution threshold value is determined based on the mean and two standard deviations of the historical contribution degree, and the advertisement features whose contribution degree is greater than or equal to the contribution threshold value are marked as high-contribution features to form a contribution feature table.

[0011] As a preferred scheme of the advertisement design auxiliary method based on artificial intelligence, wherein: including, The data collection includes the behavior types related to the advertised goods and the corresponding time stamps, and the advertisement ID data corresponding to the behaviors.

[0012] In a second aspect, the application provides an advertisement design auxiliary system based on artificial intelligence, including, The behavior matrix construction module collects the behavior data of the target user, constructs a behavior matrix, and determines the earliest time point and the diffusion parameter; The spatiotemporal distribution matrix generation module utilizes cosine similarity to measure the content correlation between behavior events, and constructs a user spatiotemporal distribution matrix in combination with the time diffusion value; The composite feature embedding module extracts advertisement features, generates composite feature embeddings, dynamically calculates behavior event weights in combination with the behavior matrix, and optimizes the spatiotemporal distribution matrix by weighting; The user interest vector generation module weights and aggregates single user behaviors based on the weighted distribution matrix, and outputs a user interest vector; The wide model feature fusion module generates a cross-feature matrix in combination with the user interest vector, the weighted distribution matrix, and the behavior time interval and content similarity, and converts the cross-feature matrix into linear feature output of a wide model; The deep model modeling module nonlinearly combines the behavior embedding result and the advertisement composite features to form an embedding vector, and utilizes a deep model to predict and output an interest matching distribution; The advertisement effectiveness quantification module generates a user interest response matrix by comparing the user interest distribution with the advertisement features, calculates the contribution degree of each advertisement feature, and quantifies the overall effectiveness of the advertisement and the distribution of the advertisement among users; Interest layering module: based on interest distribution, interest layering analysis is performed on the target user, and all advertisement features are prioritized.

[0013] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the artificial intelligence-based advertisement design assisting method according to the first aspect of the present application.

[0014] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the artificial intelligence-based advertisement design assisting method according to the first aspect of the present application.

[0015] The present application has the following beneficial effects: by introducing a local activation factor based on advertisement composite embedding, the weight of each behavior event can be dynamically associated with the difference of advertisement features, by dividing the user behavior according to the user to which the user behavior belongs, and combining the interest distribution of the time and space weighted distribution matrix, the interest distribution is further reduced to a user specific interest vector, a wide and deep model is used to generate a distribution vector based on the number of categories, and the aggregation proportion of different interest categories to user behavior and advertisement content is further clarified, by combining the overall response efficiency of the advertisement with the single user response efficiency distribution, a dynamic association between the universality evaluation of the advertisement and the individual targeted optimization can be established, by combining the interest layering and the contribution feature table, a dynamic adaptation relationship between the user characteristics and the advertisement feature priority can be established, and accurate fine-grained data support is directly provided for advertisement optimization design. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0017] Fig. 1 The flowchart of the artificial intelligence-based advertisement design assisting method in embodiment 1.

[0018] Fig. 2 The structural diagram of the artificial intelligence-based advertisement design assisting system in embodiment 1. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0020] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be appreciated that the present application can be practiced in a variety of ways beyond the specific details set forth herein, having regard to the contents of the whole patent document, and that the present application can be practiced in different but but equivalent ways without departing from the scope of the present application. It is therefore contemplated that the application that is described herein can be practiced otherwise than as specifically set forth herein.

[0021] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is independent of or mutually exclusive with other embodiments.

[0022] Embodiment 1, Reference Figs. 1-2 As a first embodiment of the present application, the embodiment provides an artificial intelligence-based advertising design assistance method, comprising the following steps: S1, data collection for target users of advertising design, construction of behavior matrix, determination of the earliest time point as reference point, standard deviation analysis to determine diffusion parameter, analysis of time diffusion value of user interest, and calculation of single time diffusion vector, use of cosine similarity to measure content correlation between behavior events, construction of space-time distribution matrix combined with time diffusion value, extraction of advertising features to form composite feature embedding, dynamic calculation of behavior event weight combined with behavior matrix, and weighting of space-time distribution matrix to obtain weighted distribution matrix, and weighted calculation of user interest vector for single user behavior; Preferably, data collection for target users of advertising design includes, Data collection includes relevant behavior categories and corresponding time stamps (such as browsing, purchasing, etc.) of advertising promotional goods, and collection of corresponding behavior advertising ID data.

[0023] Further, the time diffusion value of user interest is analyzed, and the single time diffusion vector is calculated, the cosine similarity is used to measure the content correlation between behavior events, the space-time distribution matrix is constructed combined with the time diffusion value, the advertising features are extracted to form the composite feature embedding, the behavior event weight is dynamically calculated combined with the behavior matrix, and the space-time distribution matrix is weighted to obtain the weighted distribution matrix, and the user interest vector is weighted calculated for single user behavior, including, Based on the target users of advertising design, user behavior data collection is performed, and a behavior matrix is constructed, represented as: Wherein, H represents the behavior matrix, m represents the total number of behaviors, represents the mth behavior category, represents the time stamp of the mth behavior, represents the advertising ID data of the mth behavior; The earliest time point of all user behaviors in the behavior matrix is selected as the reference point, and the standard deviation analysis is performed on all user behavior time intervals to determine the diffusion parameter suitable for the diffusion range, which is represented as: wherein, represents the time reference point, represents the behavior characteristics of the i-th row of the behavior matrix as the row vector of the i behavior, represents the time stamp of the user behavior i, represents the diffusion parameter; The time diffusion process of user interest is modeled as a quasi-ballistic diffusion, and for each behavior event in the behavior matrix, the dynamic diffusion range is calculated, and the time diffusion values are combined into a single time diffusion vector, which is represented as: wherein, represents the time diffusion value of the time, represents the diffusion vector; The cosine similarity is used to measure the content correlation between behavior events, and the time diffusion value is combined to calculate the spatio-temporal distribution value, and a spatio-temporal distribution matrix is constructed, which is represented as: wherein, represents the content correlation value of user behaviors i and j, represents the behavior characteristics of the j-th row of the behavior matrix as the row vector of the j behavior, represents the user interest spatio-temporal distribution matrix of user behaviors i and j, represents the time diffusion value of user behavior j; The pre-trained word embedding model is used to extract the advertisement text data of the advertisement content, the deep learning image coding is used to generate the image feature vector of the advertisement image, the actual category data of the advertisement is collected, and the One-Hot coding is embedded; The composite feature embedding of the target advertisement is constructed by integrating the advertisement text data, the image feature vector and the category data, the weight of the behavior event is dynamically calculated as the local activation factor combined with the behavior matrix, and the spatio-temporal distribution matrix is weighted, which is represented as: wherein, v represents the composite feature embedding, represents the local activation factor of user behavior i,​ denotes a bias for compensating the average difference of the feature value of the advertisement, denotes a weighted distribution matrix of user behaviors i and j; The user behaviors belonging to different users in the behavior matrix are divided, and the behaviors of a single user are counted. The user behavior interest vector is calculated by weighting, and is denoted as: wherein, denotes a user behavior interest vector.

[0024] Through the construction of the behavior matrix and the extraction of the time feature, combined with the dynamic modeling based on quasi-ballistic diffusion, the time sequence dynamic characteristics of the user interest are accurately captured. By introducing the timestamp and the advertisement behavior association in the behavior matrix, the earliest time point is selected as the reference point and the standard deviation analysis is performed, the user behavior time diffusion characteristics are generated, the time sequence regularity of the user behavior can be extracted mathematically, the quasi-ballistic diffusion dynamic time modeling is adopted, and the diffusion parameters of the interest are mapped to the highlight range of the user behavior. Not only can the diffusion strength of the behavior event be quantified, but also the preference difference of the user on the time distribution of different advertisement contents can be captured, so that the interest presentation is more in-depth and delicate; After introducing the time diffusion value, the time characteristics of the user behavior are converted from single time weighting to dynamic diffusion range, so that the influence of each behavior event on other behavior events can be strengthened or weakened with time, which is more in line with the real user interest propagation rule. Combined with the cosine similarity analysis of content correlation, each user behavior is dynamically modeled, and the correlation of each behavior event in time-content is quantified through the space-time distribution matrix. Compared with the traditional matrix uniform weighting method, this combined method establishes a double dimension; Through the weight calculation of the behavior event and the addition of the local activation factor, the weighting effect of the user interest response distribution in the space-time matrix is optimized. The local activation factor based on the advertisement composite embedding is introduced, so that the weight of each behavior event can be dynamically associated with the difference of the advertisement features. By dividing the user behaviors according to the user to which they belong, and combining the interest distribution calculated based on the space-time weighted distribution matrix, the interest vector specific to the user is further reduced, and the hierarchical modeling from single behavior event, local content to overall user interest is completed.

[0025] S2, using the user interest vector and the weighted distribution matrix, referring to the behavior time interval and the content similarity, generating a cross-feature matrix, and converting it into a feature fusion of a wide model, combining the behavior embedding result with the advertisement features as an embedding vector, predicting the output interest matching distribution, calculating the optimization result of the wide-deep model through nonlinear activation, and taking it as the user interest distribution; Preferably, using the user interest vector and the weighted distribution matrix, referring to the behavior time interval and the content similarity, a cross-feature matrix is generated and converted into a wide model feature fusion, the behavior embedding result is combined with the advertisement feature in a nonlinear manner as an embedding vector, the output interest matching distribution is predicted, the nonlinear activation calculation is used to calculate the optimization result of the wide-deep model, and the user interest distribution is obtained, including, Using the user interest vector and the weighted distribution matrix, combining the time interval of the access behavior and the content similarity, the behavior content and the time are processed in a unified manner, a cross-feature matrix is generated, wherein the cross-feature matrix is converted into a wide model feature fusion through weighted summation, and is represented as: wherein, represents the cross-feature of user behaviors i and j, represents the time step, and the total number of time steps after being discretized according to all time spans recorded in the user behavior matrix, represents the linear feature output of the wide model, represents the cross-feature comprehensive result of user behavior i, which is obtained by weighting and aggregating all cross-features of user behavior i, represents the weight of user behavior i in the behavior matrix, represents the data frequency of user behavior i, represents the total data frequency in the behavior matrix; The standard sequence embedding framework based on the user behavior sequence is adopted to capture the dependence relationship between the behavior feature and the time diffusion value, the Transformer is used to model the time sequence, and the time diffusion characteristic is integrated into the sequence embedding. The behavior embedding result is combined with the advertisement feature in a nonlinear manner as an embedding vector, and is represented as: wherein, represents the dependence relationship representation vector of user behavior i, represents the dependence relationship representation of the time dependence of the behavior sequence to the current behavior, which is based on the output of the Transformer model, represents a nonlinear activation function, represents the high-dimensional vector expression of the fusion of user behavior i and the advertisement feature, as an embedding vector; The deep model is constructed by using a fully connected feedforward neural network, and is pre-trained according to historical data; the embedding vector, the composite feature embedding and the user behavior interest vector are normalized and composed into an input vector, and the output interest matching distribution is predicted; The interest matching distribution is combined with the wide model linear feature output, the optimization result of the wide deep model is calculated through nonlinear activation, and is used as the user interest distribution, wherein the embedding vector and the wide model linear feature output are projected to the same dimension d through the construction of a fully connected layer, d is the number of user interest categories, the generated user interest distribution is calculated through joint calculation, the adaptability of the advertisement system to user interest features is improved, and is expressed as: Wherein, represents the interest matching distribution, represents the weight of the deep model, represents the bias of the deep model, represents the activation function, represents the user interest distribution, and represents the overall interest of the user to the advertisement.

[0026] The linear feature fusion of the cross feature matrix in the wide model through weighted summation can effectively measure the specific preference of the user to the advertisement content under the dynamic time span, especially in the case that the behavior data has time heterogeneity, the rich behavior personality of the user can be accurately captured, and the advertisement optimization ability based on the linear feature is further improved; The time diffusion characteristic dynamically quantifies the time sensitivity of the user behavior, and maps this weight to the global context of the user behavior through the sequence embedding vector output by the Transformer, compared with the analysis of the behavior frequency or the static time window in the traditional method, this sequence embedding framework can capture the complex changes of the behavior timing pattern in a more accurate way, thereby providing a higher-dimensional and more information-containing behavior vector description, laying a semantic fusion foundation for subsequent nonlinear feature modeling; Since the advertisement itself is composed of text, image and category multi-modal features, and the behavior embedding contains the deep mixed relationship of time and content, the nonlinear combination can learn the complex influence path between the time evolution of the user interest and the advertisement features, break through the behavior-advertisement interaction limitation under the single mode and linear feature representation, and thus provide highly refined semantic features for constructing the real user interest; Through the construction of a fully connected feedforward neural network (deep model), the embedding vector, the user behavior interest and the advertisement composite feature vector are integrated, the data features of different sources are finely fused, and the adaptability to complex behavior patterns is optimized through the deep learning ability of the model; Through the wide-deep collaborative mechanism, the model can not only accurately quantify the actual interest distribution of the user, but also consider the behavior logic, the content characteristics of the advertisement, and the multi-time scale change, thereby effectively improving the adaptability of the advertisement system to the real personal user interest preference, satisfying the expression of the general behavior mode and flexibly capturing the personalized behavior characteristics of the user, and enhancing the overall advertisement delivery effect; The joint wide-deep model generates a distribution vector based on the number of categories for the user interest graph, further clarifies the aggregation proportion of different interest categories to the user behavior and the advertisement content, and makes the overall interest of the user to the advertisement not only a set of predicted values, but also a distributed expression with different interest categories as semantic dimensions, which can better support the multi-dimensional refinement operation of the advertisement target audience and the accurate guidance of the content optimization direction.

[0027] S3, by comparing the user interest distribution and the advertisement characteristics one by one, a user interest response matrix is generated, the contribution degree of each advertisement characteristic is analyzed, the response degree of the user interest distribution on different advertisement characteristics is quantified, and the overall effectiveness of the advertisement is quantified through a comprehensive index, and the distribution of the advertisement effectiveness on different users is further quantified; Preferably, the user interest response matrix is generated, the contribution degree of each advertisement characteristic is analyzed, the response degree of the user interest distribution on different advertisement characteristics is quantified, and the overall effectiveness of the advertisement is quantified through a comprehensive index, and the distribution of the advertisement effectiveness on different users is further quantified, including, The user interest distribution is used to represent the interest intensity of the user to the advertisement, which is compared with the advertisement characteristics one by one to generate a user interest response matrix, the advertisement characteristics (columns of the user interest response matrix) of the user interest response matrix are normalized, the contribution degree of each advertisement characteristic is analyzed, and is expressed as: Among them, represents the user interest response of the user q to the advertisement characteristic g, represents the user interest distribution of the user q, represents the gth feature value of the composite feature embedding, represents the contribution proportion of the gth advertisement characteristic to the total interest response, Q represents the total number of users, and G represents the total number of features of the composite feature embedding; The contribution value of the advertisement characteristic is used to quantify the response degree of the user interest distribution on different advertisement characteristics, and the overall effectiveness of the advertisement is quantified through a comprehensive index, and the distribution of the advertisement effectiveness on different users is further quantified in addition to the overall effectiveness of a single advertisement, and is expressed as: Among them, an overall response effectiveness of the advertisement, an interest response effectiveness of the advertisement to the user q.

[0028] By fusing the spatiotemporal characteristics, content correlation and weighted dynamic optimization of user behaviors through the user interest distribution, and combining with the one-by-one comparison of the advertisement features, the personalized needs of the user can be accurately restored in the feature dimension of the multi-modal advertisement content, thereby providing specific response data basis for subsequent advertisement optimization design; By normalizing the advertisement feature columns in the user interest response matrix, the contribution degree of each advertisement feature is calculated, the feature dominant effect caused by the too large or too small of a single advertisement feature value is effectively avoided, the internal strong and weak relationship between the user interest distribution and the advertisement features is standardized, and the influence degree of different advertisement fine-grained features on the user interest is accurately quantified; Through the advertisement feature contribution value, the response degree of the user interest on different advertisement features is quantified, the source of the attraction of the advertisement to the target user can be resolved from the feature level, not only the utility distribution of the advertisement on a single user can be presented, but also the different feature preferences of the advertisement in different user groups can be compared vertically; The quantification of the overall response effectiveness of the advertisement significantly optimizes the evaluation of the advertisement effect, the value is no longer constrained by the traditional indicators such as click rate, but becomes the accurate mapping based on the mutual matching of the user interest and the advertisement features, which can reflect the universality and attraction boundary of the advertisement across user groups, and provides an explicit and reliable overall evaluation for predicting the performance of the advertisement; Further quantifying the distribution of the advertisement effectiveness on different users can help the advertisement evaluation to establish a bridge between the group features and the individual features, thereby targetedly subdividing the target user group and providing personalized delivery strategies according to the quantification results of the individual effectiveness; By combining the overall response effectiveness of the advertisement with the single-user response effectiveness distribution, a dynamic association between the universality evaluation of the advertisement and the individual targeted optimization can be established, in the advertisement with high overall response effectiveness, the effectiveness distribution results of the user personalized preferences can be further utilized to provide a multi-level feedback mechanism: while ensuring that the general features are dominant in the advertisement design, the high-weighted user preference features are optimized.

[0029] S4, performing interest stratification analysis on the target user, and performing priority sorting on all features of the advertisement to form a contribution feature table; Preferably, the interest stratification analysis on the target user comprises, According to the interest response effectiveness data of the advertisement to the user, the interest threshold is defined based on the historical data, the interest threshold is determined by the mean and two standard deviations of the historical interest effectiveness, the user with the interest response effectiveness greater than or equal to the interest threshold is marked as a high-interest user, and the user with the interest response effectiveness less than the interest threshold is marked as a low-interest user, and the interest stratification analysis is performed on the target user.

[0030] The distinction between high-interest users and low-interest users generated by the interest layering mechanism can help the advertising delivery system further optimize resource allocation. For example, the accurate identification of high-interest users enables the advertising to concentrate on tilting the resource allocation to these users, while low-interest users can implement more attractive content stimulation mechanisms, thereby optimizing the delivery effect of the advertising for different interest levels. At the same time, the results of interest layering provide distribution data support for the effectiveness evaluation of the advertising to the users, which not only ensures that the advertising system can efficiently cover the user group more sensitive to the advertising, but also avoids wasting resources on groups with large differences in effect.

[0031] Further, all features of the advertising are prioritized to form a contribution feature table, including, Based on the contribution degree of the advertising features, all features of the advertising are prioritized, and the contribution threshold is determined based on the mean and two standard deviations of the historical contribution degree. The advertising features with a contribution degree greater than or equal to the contribution threshold are marked as high-contribution features to form a contribution feature table.

[0032] The construction of the contribution feature table not only serves as a simple sorting of the feature influence, but also directly provides accurate fine-grained data support for the advertising optimization design. For example, by marking the high-contribution features to highlight in the advertising design, the user's attention can be concentrated on the content part that has a greater impact on interest matching, significantly reducing the resource investment in low-contribution details, so that the advertising design can improve the user interaction rate and attractiveness in a more efficient way. Through the combination of interest layering and the contribution feature table, a dynamic adaptation relationship between user characteristics and advertising feature priorities can be established. Specifically, high-interest users can be matched with high-contribution features in advertising delivery, ensuring that resource allocation is concentrated on the efficient combination of the two; low-interest users can strictly exclude low-contribution features in advertising design, and use enhanced design of high-contribution features to try to improve interactive interest, in a flexible way to improve the reach effect of advertising to low-interest users.

[0033] The embodiment also provides an advertising design auxiliary system based on artificial intelligence, comprising, a behavior matrix construction module that collects behavior data of target users, constructs a behavior matrix, and determines an earliest time point and a diffusion parameter; a spatiotemporal distribution matrix generation module that uses cosine similarity to measure the content correlation between behavior events and constructs a user spatiotemporal distribution matrix in combination with a time diffusion value; a composite feature embedding module that extracts advertising features, generates composite feature embeddings, dynamically calculates behavior event weights in combination with the behavior matrix, and optimizes the spatiotemporal distribution matrix by weighting; A user interest vector generation module: based on the weighted distribution matrix, the single user behavior is weighted and aggregated, and the user interest vector is output; A wide model feature fusion module: combining the user interest vector, the weighted distribution matrix, and the behavior time interval and content similarity, generating a cross-feature matrix and converting it into a linear feature output of the wide model; A deep model modeling module: the behavior embedding result and the advertisement composite feature are combined non-linearly to form an embedding vector, and a deep model is used to predict the interest matching distribution; An advertisement performance quantification module: by comparing the user interest distribution with the advertisement features, a user interest response matrix is generated, the contribution of each advertisement feature is calculated, and the overall performance of the advertisement and its distribution among users are quantified; An interest layering module: based on the interest distribution, the target user is analyzed for interest layering, and all advertisement features are prioritized.

[0034] The embodiment also provides a computer device suitable for the case of the advertisement design auxiliary method based on artificial intelligence, which comprises a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the advertisement design auxiliary method based on artificial intelligence proposed in the above embodiment.

[0035] The computer device can be a terminal, which comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. In addition, an external keyboard, touchpad or mouse can also be used.

[0036] The embodiment also provides a storage medium on which a computer program is stored, the program being executed by a processor to implement the method for assisting advertisement design based on artificial intelligence proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk, or an optical disk.

[0037] To sum up, by introducing the local activation factor based on the advertisement composite embedding, the weight of each behavior event can be dynamically associated with the difference of the advertisement features, the user behaviors are divided according to the users, the interest distribution is counted by combining the time-space weighted distribution matrix, and is further reduced to the user-specific interest vector, the distribution vector based on the number of categories is generated by combining the wide-deep model to generate the user interest graph, the aggregation proportion of different interest categories to the user behaviors and the advertisement content is further clarified, the overall response efficiency of the advertisement and the single-user response efficiency distribution are combined for use, the dynamic association between the universality evaluation of the advertisement and the individual targeted optimization can be established, by combining the interest layering and the contribution feature table, the dynamic adaptation relationship between the user characteristics and the advertisement feature priority can be established, and accurate fine-grained data support is directly provided for the advertisement optimization design.

[0038] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. An artificial intelligence-based advertisement design assistance method, characterized by, The method comprises the following steps: Data collection is performed for target users of an advertisement, a behavior matrix is constructed, the earliest time point is determined as a reference point, standard deviation analysis is performed to determine diffusion parameters, time diffusion values of user interest are analyzed, a single time diffusion vector is calculated, content correlation between behavior events is measured using cosine similarity, a space-time distribution matrix is constructed by combining time diffusion values, composite feature embedding is extracted from advertisement features, the weight of a behavior event is dynamically calculated based on the behavior matrix, and a weighted distribution matrix is obtained by weighting the space-time distribution matrix, and a user interest vector is calculated by weighting the behavior of a single user; Using the user interest vector and the weighted distribution matrix, referring to the behavior time interval and the content similarity, a cross-feature matrix is generated, and is converted into a feature fusion of a wide model. The behavior embedding result and the advertisement features are combined non-linearly as an embedding vector, an interest matching distribution is predicted, the optimization result of the wide-deep model is calculated through non-linear activation, and is used as the user interest distribution; By comparing the user interest distribution and the advertisement features one by one, a user interest response matrix is generated, the contribution degree of each advertisement feature is analyzed, the response degree of the user interest distribution on different advertisement features is quantified, and the overall performance of the advertisement is quantified through a comprehensive index, and the distribution of the advertisement performance on different users is further quantified; The interest of the target user is analyzed, and the priority of all features of the advertisement is sorted to form a contribution feature table. 2.The artificial intelligence-based advertisement design assistance method of claim 1, wherein: The analysis of the time diffusion values of the user interest and the calculation of the single time diffusion vector, the measurement of the content correlation between the behavior events using the cosine similarity, the construction of the space-time distribution matrix by combining the time diffusion values, the extraction of the composite feature embedding from the advertisement features, the dynamic calculation of the weight of the behavior event based on the behavior matrix, the weighting of the space-time distribution matrix to obtain the weighted distribution matrix, and the calculation of the user interest vector by weighting the behavior of a single user, comprise: Based on the target user of the advertisement, user behavior data is collected, and a behavior matrix is constructed; The earliest time point of all user behaviors in the behavior matrix is selected as a reference point, and the standard deviation analysis is performed on all user behavior time intervals, and then the diffusion parameters suitable for the diffusion range are determined; The time diffusion process of the user interest is modeled as a quasi-ballistic diffusion, the dynamic diffusion range of each behavior event in the behavior matrix is calculated, and the time diffusion values are combined into a single time diffusion vector; The content correlation between the behavior events is measured using the cosine similarity, and the space-time distribution values are calculated by combining the time diffusion values, and a space-time distribution matrix is constructed; The pre-trained word embedding model is used to extract the advertisement text data of the advertisement content, the deep learning image coding is used to generate the image feature vector of the advertisement image, the actual category data of the advertisement is collected, and the One-Hot coding and embedding are performed; The composite feature embedding of the target advertisement is constructed by integrating the advertisement text data, the image feature vector and the category data, the weight of the behavior event is dynamically calculated as a local activation factor based on the behavior matrix, and the space-time distribution matrix is weighted; Different behaviors of the users in the behavior matrix are divided, and the behaviors of a single user are counted to calculate the user behavior interest vector. 3.The artificial intelligence-based advertisement design assistance method of claim 2, wherein: The use of user interest vector and weighted distribution matrix, reference behavior time interval and content similarity, generate cross feature matrix, and convert to wide model feature fusion, nonlinear combination of behavior embedding result and advertisement feature as embedding vector, predict output interest matching distribution, calculate the optimization result of wide deep model through nonlinear activation, and take it as user interest distribution, including, Using user interest vector and weighted distribution matrix, combining the time interval of access behavior and content similarity, processing behavior content and time together, generating cross feature matrix, wherein the cross feature matrix is converted into wide model feature fusion by weighted summation; Using the standard sequence embedding framework based on user behavior sequence, using Transformer to model the time sequence, and integrating the time diffusion characteristics into the sequence embedding, nonlinearly combining the behavior embedding result and the advertisement feature as the embedding vector; Using a fully connected feedforward neural network to build a deep model, and pre-training according to historical data, normalizing and combining the embedding vector, composite feature embedding, and user behavior interest vector to form an input vector, and predicting the output interest matching distribution; Combine the interest matching distribution with the linear feature output of the wide model, calculate the optimization result of the wide deep model through nonlinear activation, and take it as the user interest distribution.

4. The artificial intelligence-based advertisement design assistance method of claim 3, wherein: The generation of user interest response matrix, the contribution degree of each advertisement feature is analyzed, the response degree of user interest distribution on different advertisement features is quantified, and the overall performance of the advertisement is quantified through comprehensive indicators, and the distribution of advertisement performance on different users is further quantified, including, Using user interest distribution to represent the interest intensity of users to advertisements, comparing them one by one with advertisement features, generating user interest response matrix, normalizing the advertisement features of user interest response matrix, and analyzing the contribution degree of each advertisement feature; Using the contribution value of the advertisement feature, quantifying the response degree of the user interest distribution on different advertisement features, and quantifying the overall performance of the advertisement through comprehensive indicators, and further quantifying the distribution of advertisement performance on different users in addition to the overall performance of a single advertisement. 5.The artificial intelligence-based advertisement design assistance method of claim 4, wherein: The interest layer analysis of the target user, Including, Based on the historical data, the interest threshold is defined based on the historical data, the mean and two standard deviations of the historical interest performance are determined, the users with interest response performance greater than or equal to the interest threshold are marked as high-interest users, and the users with interest response performance less than the interest threshold are marked as low-interest users, and the target user is analyzed by interest layer. 6.The artificial intelligence-based advertisement design assistance method of claim 5, wherein: The priority of all features of the advertisement is sorted to form a contribution feature table, including, Based on the contribution degree of the advertisement feature, the priority of all features of the advertisement is sorted, the contribution threshold is determined based on the mean and two standard deviations of the historical contribution degree, the advertisement features with contribution degree greater than or equal to the contribution threshold are marked as high-contribution features, and a contribution feature table is formed. 7.The artificial intelligence-based advertisement design assistance method of claim 2, wherein: The data collection of the target user designed for the advertisement, including, The data collection includes the behavior types related to the advertisement promotion goods and the corresponding time stamps, and the advertisement ID data of the corresponding behavior.

8. An artificial intelligence-based advertisement design assistance system based on any one of the artificial intelligence-based advertisement design assistance methods according to claims 1 to 7, characterized by: Including, The behavior matrix construction module: collects the behavior data of the target user, constructs a behavior matrix, and determines the earliest time point and diffusion parameters; The space-time distribution matrix generation module: uses cosine similarity to measure the content correlation between behavior events, and constructs a user space-time distribution matrix combined with the time diffusion value; The composite feature embedding module: extracts advertisement features, generates composite feature embeddings, dynamically calculates behavior event weights based on the behavior matrix, and optimizes the space-time distribution matrix; The user interest vector generation module: weights and aggregates single user behavior based on the weighted distribution matrix, and outputs a user interest vector; The wide model feature fusion module: generates a cross-feature matrix and converts it into a linear feature output of the wide model, combining the user interest vector, the weighted distribution matrix, and the behavior time interval and content similarity; The deep model modeling module: combines the behavior embedding result and the advertisement composite feature in a nonlinear manner to form an embedding vector, and uses a deep model to predict an interest matching distribution; The advertisement effectiveness quantification module: generates a user interest response matrix by comparing the user interest distribution with the advertisement features, calculates the contribution of each advertisement feature, and quantifies the overall effectiveness of the advertisement and its distribution among users; The interest layering module: performs interest layering analysis on the target user based on the interest distribution, and prioritizes all advertisement features. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the advertisement design auxiliary method based on artificial intelligence in any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the advertisement design auxiliary method based on artificial intelligence in any one of claims 1-7.

Images