Advertisement putting method and device and storage medium

Abstract

The embodiment of the invention provides an advertisement putting method and device and a storage medium, and the method comprises the steps: obtaining a portrait feature vector of a client, and retrieving a plurality of target advertisement feature vectors, the similarity of which to the portrait feature vector meets the requirement, from a vector database; and according to a corresponding relationship between the advertisement feature vectors and the to-be-delivered advertisements, determining the to-be-delivered advertisements corresponding to the plurality of target advertisement feature vectors. And according to the advertisement to be put and the portrait feature vector of the client, generating a target prompt text corresponding to the large language model. And inputting the target prompt text into the large language model to obtain a target advertisement matched with the client. According to the scheme, the advertisement is pointedly put to the client according to the portrait information of the client through two times of screening, so that the advertisement putting accuracy and conversion rate in the application program can be improved, and the user experience can also be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an advertisement delivery method, device and storage medium. Background Art

[0002] With the development of the Internet and mobile terminals, the public's attention has shifted to mobile terminals. Currently, a large number of advertisements have begun to be placed in third-party applications on such mobile devices. Currently, before placing advertisements in applications, historical data corresponding to the client is usually collected first, user interests and hobbies are analyzed, and then advertisements related to the user's interests are pushed to the user. However, this advertising delivery solution has a vague control over the user's potential needs, resulting in low accuracy in advertising delivery, which not only affects the advertising delivery effect, but also affects the user's experience of using the application.

[0003] Therefore, how to accurately place advertisements has become an urgent problem to be solved. Summary of the invention

[0004] The embodiments of the present invention provide an advertisement delivery method, device and storage medium, which are used to deliver advertisements to a client in a targeted manner, improve the accuracy of advertisement delivery within an application program, and thus enhance user experience.

[0005] In a first aspect, an embodiment of the present invention provides an advertisement delivery method, the method comprising:

[0006] If it is detected that the client starts the target application, a portrait feature vector of the client is obtained;

[0007] Retrieving a plurality of target advertisement feature vectors whose similarity with the portrait feature vector meets the requirement from a vector database, wherein the vector database stores advertisement feature vectors corresponding to a plurality of advertisements to be placed;

[0008] Determining the advertisements to be placed corresponding to each of the plurality of target advertisement feature vectors according to the corresponding relationship between the advertisement feature vectors and the advertisements to be placed;

[0009] Generate a target prompt text corresponding to the large language model according to the advertisement to be placed and the portrait feature vector, wherein the target prompt text is used to instruct the large language model to filter out a target advertisement from a plurality of advertisements to be placed corresponding to each of the plurality of target advertisement feature vectors;

[0010] The target prompt text is input into a large language model to obtain a target advertisement matching the client.

[0011] In a second aspect, an embodiment of the present invention provides an advertisement delivery device, the device comprising:

[0012] A monitoring module, for obtaining a portrait feature vector of the client if it is detected that the client starts a target application;

[0013] A retrieval module, used to retrieve a plurality of target advertisement feature vectors whose similarity with the portrait feature vector meets the requirements from a vector database, wherein the vector database stores advertisement feature vectors corresponding to a plurality of advertisements to be placed;

[0014] A determination module, configured to determine the advertisements to be placed corresponding to each of the plurality of target advertisement feature vectors according to the correspondence between the advertisement feature vectors and the advertisements to be placed;

[0015] A generating module, configured to generate a target prompt text corresponding to a large language model according to the advertisement to be placed and the portrait feature vector, wherein the target prompt text is used to instruct the large language model to filter out a target advertisement from a plurality of advertisements to be placed corresponding to each of the plurality of target advertisement feature vectors;

[0016] The processing module is used to input the target prompt text into a large language model to obtain a target advertisement matching the client.

[0017] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the advertising delivery method as described in the first aspect.

[0018] In a fourth aspect, an embodiment of the present invention provides a non-transitory machine-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of an electronic device, the processor can at least implement the advertising delivery method as described in the first aspect.

[0019] In the advertising delivery scheme provided by the embodiment of the present invention, during the period when the client uses the application, the target advertisement that the user is interested in is matched for the client, so as to achieve targeted delivery of the target advertisement for the client. Specifically, it is monitored whether the client starts the target application. If it is monitored that the client starts the target application, the portrait feature vector of the client is obtained, and multiple target advertisement feature vectors that meet the requirements of similarity with the portrait feature vector are retrieved from the vector database. Among them, the vector database stores multiple advertisement feature vectors corresponding to the advertisements to be delivered. Then, according to the correspondence between the advertisement feature vector and the advertisements to be delivered, the advertisements to be delivered corresponding to each of the multiple target advertisement feature vectors are determined. According to the advertisements to be delivered corresponding to each of the multiple target advertisement feature vectors and the portrait feature vector of the client, the target prompt text corresponding to the large language model is generated. Among them, the target prompt text is used to instruct the large language model to filter out the target advertisement from the multiple advertisements to be delivered corresponding to each of the multiple target advertisement feature vectors. Finally, the target prompt text is input into the large language model to obtain the target advertisement that matches the client.

[0020] In the above scheme, by comparing the portrait feature vector with the advertisement feature vectors corresponding to multiple advertisements to be placed in the vector database, multiple advertisements to be placed that match the client portrait are screened out, and based on the multiple advertisements to be placed and the portrait feature vectors, a target prompt text is generated, and the target prompt text is input into the large language model so that the large language model can better understand the semantic information, so as to screen out the target advertisement with the highest degree of match with the client from the multiple advertisements to be placed, that is, through two screenings, it is achieved to deliver advertisements to the client in a targeted manner according to the client's feature information, which can not only improve the accuracy and conversion rate of advertising delivery within the application, but also improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 A schematic diagram of an application scenario of an advertisement delivery method provided by an embodiment of the present invention;

[0023] Figure 2 A flowchart of an advertisement delivery method provided by an embodiment of the present invention;

[0024] Figure 3 A flowchart of determining a client portrait feature vector provided by an embodiment of the present invention;

[0025] Figure 4 A schematic diagram of a flow chart of another advertisement delivery method provided by an embodiment of the present invention;

[0026] Figure 5 A flowchart of generating a target prompt text corresponding to a large language model provided by an embodiment of the present invention;

[0027] Figure 6 A schematic diagram of the structure of an advertisement delivery device provided by an embodiment of the present invention;

[0028] Figure 7 A schematic diagram of the structure of an electronic device provided in this embodiment. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the step timing in the following method embodiments is only an example, not a strict limitation.

[0030] It should be noted that the user data (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0031] As the number of mobile terminal users continues to expand, advertisers' demand for mobile advertising continues to grow. Currently, a large number of advertisements are being placed in third-party applications on such mobile devices. The current advertising delivery solution mainly collects historical data corresponding to the client, analyzes user interests and hobbies, and then pushes advertisements related to the user's interests to the user. However, this advertising delivery solution has a vague control over the user's potential needs, resulting in low accuracy in advertising delivery, which not only affects the advertising delivery effect, but also affects the user's experience of using the application.

[0032] Therefore, in order to solve the above technical problems, the present invention provides a new advertising delivery method, which uses big model technology to extract feature information corresponding to each advertisement to be delivered in the advertising data set and the portrait information of the client, and compares the feature information corresponding to each advertisement to be delivered with the portrait information of the client, so as to retrieve multiple advertisements to be delivered that match the portrait information of the client from the advertising data set, and uses the AI ​​big language model in combination with the portrait information of the client to screen out the target advertisement with the highest degree of match with the user from the multiple matching advertisements to be delivered, so as to automatically deliver the target advertisement while the client is using the application. This can not only improve the accuracy and conversion rate of advertising delivery, but also enhance the user experience.

[0033] In the following, some embodiments of the present invention are described in detail in conjunction with the accompanying drawings. In the case where there is no conflict between the embodiments, the following embodiments and features in the embodiments can be combined with each other.

[0034] Figure 1 A schematic diagram of an application scenario of an advertisement delivery method provided by an embodiment of the present invention; Figure 1 As shown, the execution subject of the method may be an advertisement delivery device, which may be implemented as software or a combination of software and hardware, and the advertisement delivery device may be communicatively connected with a client.

[0035] The client may refer to a user device with a target application installed. The target application may be any type of application, and the embodiments of the present invention do not limit this. The advertisement delivery device may automatically deliver advertisement services on the client, and dynamically adjust the advertisements delivered in the target application in the client according to the actual needs and interests of the user, so as to more accurately deliver the advertisements that the user is interested in, which can not only improve the advertisement conversion rate but also enhance the user experience.

[0036] Specifically, the processing process of the advertisement delivery device for advertisement delivery on the client side may be: the advertisement delivery device may regularly collect basic user data corresponding to the client side, historical operation data corresponding to the user when using the target application, and user browsing history, etc. Based on the various information corresponding to the client side collected, the user portrait information corresponding to the client side is generated. In order to have a deeper understanding of the user's interests, preferences, behavior patterns and other information, and thus to deliver advertisements to the user more accurately, after obtaining the user portrait information, the user portrait information may be subjected to feature extraction processing to obtain the portrait feature vector corresponding to the client side. In addition, the collected user data and portrait feature vector may be stored in a preset storage area, so that the portrait feature vector corresponding to the client side can be directly obtained from the preset storage area later.

[0037] Furthermore, the advertisement delivery device may also periodically collect advertisements to be delivered, and update the data to be delivered to the advertisement data set. In order to better match the target advertisement with the highest similarity to the user's interests and hobbies for the client, when delivering the target advertisement to the client, each advertisement to be delivered in the advertisement data set may be analyzed first to extract the advertisement feature vector corresponding to each advertisement to be delivered, and store the advertisement feature vector corresponding to each advertisement to be delivered in the vector database, so that when delivering advertisements to the client later, the advertisement feature vector corresponding to each advertisement to be delivered may be directly obtained from the vector database, thereby improving processing efficiency.

[0038] In addition, during specific implementation, user data and the advertisements to be delivered in the advertisement data set may be updated regularly to timely adapt to user changes and ensure the effectiveness of advertisement delivery.

[0039] After completing the above processing, the startup status of the target application in the client is monitored. If the client is detected to have started the target application, the target advertisement is matched for the client and placed in the target application corresponding to the client. In this way, when the user starts the target application through the client, the target advertisement with the highest similarity to the user's interests and hobbies can be automatically placed while the user is using the target application. This eliminates the need for manual operation by advertising operators and is more intelligent and efficient.

[0040] The specific implementation process of matching target ads for clients includes: obtaining the client's portrait feature vector, retrieving multiple target ad feature vectors whose similarity with the portrait feature vector meets the requirements from the vector database. And according to the correspondence between the ad feature vector and the ad to be placed, the ad to be placed corresponding to each of the multiple target ad feature vectors is determined. The ad feature vector can be regarded as a vector representation of the ad to be placed, which contains key information in the ad to be placed.

[0041] Next, according to the multiple advertisements to be placed and the portrait feature vectors corresponding to the multiple target advertisement feature vectors, a target prompt text corresponding to the large language model is generated. The target prompt text is used to instruct the large language model to select the target advertisement from the multiple advertisements to be placed corresponding to the multiple target advertisement feature vectors.

[0042] Then, the target prompt text is input into the large language model to obtain a target advertisement that matches the client, and the target advertisement is placed in the target application. When placing the target advertisement in the target application, the placement location, placement method, and placement time corresponding to the target advertisement can be determined first. In specific implementation, the placement location, placement method, and placement time corresponding to the target advertisement can be determined based on user behavior data, user usage habits, and data corresponding to the advertisement clicked by the user.

[0043] The technical solution provided in this embodiment compares the advertisement feature vectors corresponding to each advertisement to be placed with the portrait feature vector of the client, so as to retrieve multiple advertisements to be placed that match the portrait feature vector of the client from the vector database, and uses the AI ​​big language model in combination with the portrait feature vector of the client to screen out the target advertisement with the highest degree of match with the user from the multiple matching advertisements to be placed, so as to automatically place the target advertisement while the client is using the application. This can not only improve the accuracy and conversion rate of advertisement placement, but also enhance the user experience.

[0044] The above embodiment describes the advertisement delivery process in a specific application scenario. In order to more clearly understand the above advertisement delivery process, the process is described in detail in conjunction with the following embodiment.

[0045] Figure 2 A flowchart of an advertisement delivery method provided by an embodiment of the present invention is shown in FIG. Figure 2 As shown, the method comprises the following steps:

[0046] 201. If it is detected that the client starts the target application, a portrait feature vector of the client is obtained.

[0047] 202. Retrieve a plurality of first target advertisement feature vectors whose similarity with the portrait feature vector meets the requirement from a vector database, wherein the vector database stores advertisement feature vectors corresponding to a plurality of advertisements to be placed.

[0048] 203. Determine the advertisements to be placed corresponding to the multiple target advertisement feature vectors according to the corresponding relationship between the advertisement feature vectors and the advertisements to be placed.

[0049] 204. Generate a target prompt text corresponding to the large language model based on the multiple advertisements to be placed and the portrait feature vectors. The target prompt text includes instructions for instructing the large language model to filter out the target advertisement from the multiple advertisements to be placed corresponding to the multiple target advertisement feature vectors.

[0050] 205. Input the target prompt text into the large language model to obtain a target advertisement that matches the client.

[0051] In the advertising delivery solution provided by the embodiment of the present invention, when the client uses the target application, the target advertisement that best matches the client's interests and hobbies can be automatically delivered in the target application. In specific implementation, it can be monitored whether the target application in the client is started. If it is monitored that the client starts the target application, the client's portrait feature vector is obtained. Among them, the client's portrait feature vector can comprehensively reflect the user's characteristics from multiple dimensions, providing strong data support for matching the user with target advertisements.

[0052] In the specific implementation, the user data corresponding to the client can be collected first. Among them, the user data includes the user's basic information, user behavior data, interest preferences, usage habits, etc. Basic information, such as age, gender, geographic location, occupation, etc. User behavior data, such as frequency of use, access time, interaction method, etc. Interest preferences such as favorite content types, browsing habits, etc. Then, the collected user data is pre-processed by cleaning, deduplication, format conversion, etc. to ensure the accuracy and availability of the data. Then, the processed data is subjected to feature extraction processing to obtain user portrait information. Next, the user portrait information corresponding to the client is quantified to obtain the client's portrait feature vector. And the client's portrait feature vector is stored in a preset storage area. Among them, the portrait feature vector can more comprehensively reflect the user's characteristics, so that when matching advertisements, the target advertisements that match the user can be more accurately screened out.

[0053] Then, when the client is detected to start the target application, the client's portrait feature vector can be directly read from the preset storage area. Then, by analyzing the client's portrait feature vector and multiple advertising feature vectors in the vector data, multiple target advertising feature vectors that meet the requirements of similarity with the portrait feature vector are retrieved from the vector database.

[0054] The vector database stores multiple advertisement feature vectors corresponding to the advertisements to be placed. The advertisement feature vectors can reflect the feature information corresponding to the advertisements to be placed from multiple dimensions, so that when searching from the vector database, multiple target advertisement feature vectors that meet the requirements of similarity with the user portrait feature vector can be retrieved more accurately.

[0055] In addition, in practical applications, the advertising feature vectors corresponding to each advertisement to be placed can be determined in advance, and the advertising feature vectors corresponding to each advertisement to be placed can be stored in the vector database. Then, when screening target advertisements for the client, the advertising feature vectors corresponding to each advertisement to be placed in the vector database can be directly used for similarity matching processing to screen out multiple target feature vectors whose similarity with the client's portrait feature vector meets the requirements.

[0056] Regarding the implementation method of determining the advertisement feature vector corresponding to each advertisement to be placed, in an optional embodiment, the advertisement type corresponding to each advertisement to be placed in the advertisement data set may be determined first, and based on the advertisement type corresponding to each advertisement to be placed, a pre-trained model may be used to perform feature extraction processing on the advertisement to be placed to obtain the advertisement feature vector corresponding to each advertisement to be placed. In addition, after determining the advertisement feature vector corresponding to each advertisement to be placed, the advertisement feature vector corresponding to each advertisement to be placed may be stored in a vector database.

[0057] Since the vector database stores the advertisement feature vectors corresponding to each advertisement to be placed, in order to facilitate the subsequent rapid screening of advertisements to be placed that match the user, when the advertisement feature vectors corresponding to each advertisement to be placed are stored in the vector database, the corresponding relationship between each advertisement feature vector and the corresponding advertisement to be placed can be established at the same time. Among them, the corresponding relationship between the advertisement feature vector and the corresponding advertisement to be placed can be stored in a mapping table, and the advertisement to be placed corresponding to each advertisement feature vector can be obtained by traversing the mapping table. In other words, before screening the target advertisements that match the client, a vector database can be established first, the advertisement feature vectors corresponding to each advertisement to be placed can be determined, and the advertisement feature vectors can be stored in the vector database.

[0058] In practical applications, after completing the above operations, the advertisement feature vectors corresponding to each advertisement to be placed and the client's portrait feature vectors can be directly stored in the vector database to retrieve the target advertisement feature vector whose similarity with the client's portrait feature vector meets the requirements.

[0059] Regarding the specific implementation method of retrieving multiple target advertising feature vectors whose similarity with the client's portrait feature vector meets the requirements from the vector database, in an optional embodiment, a similarity threshold can be set according to actual needs. When the similarity between the advertising feature vector in the vector database and the client's portrait feature vector meets the preset similarity threshold, the advertising feature vector is determined as a target advertising feature vector whose similarity with the client's portrait feature vector meets the requirements.

[0060] Specifically, the similarity between the client's portrait feature vector and each advertising feature vector in the vector database is determined respectively. When the similarity between a certain advertising feature vector and the client's portrait feature vector meets a preset similarity threshold, the advertising feature vector is determined as a target advertising feature vector whose similarity with the client's portrait feature vector meets preset requirements.

[0061] Regarding the specific implementation method of retrieving multiple target advertising feature vectors that meet the requirements of similarity with the client's portrait feature vector from a vector database, in another optional embodiment, the similarity between the client's portrait feature vector and each advertising feature vector in the vector database can be determined respectively, and the multiple advertising feature vectors can be sorted in order of similarity from large to small, and a preset number of multiple target advertising feature vectors with a high ranking can be selected. Among them, the client's portrait feature vector includes an interest feature vector, so it is also possible to directly determine the similarity between the client's interest feature vector and each advertising feature vector in the vector database, sort the multiple advertising feature vectors in order of similarity from large to small, and select a preset number of multiple target advertising feature vectors with a high ranking.

[0062] In practical applications, the similarity between each advertisement feature vector and the client's portrait feature vector can be determined by calculating the cosine similarity, Euclidean distance, etc. between each advertisement feature vector in the vector database and the client's portrait feature vector. Alternatively, the similarity between each advertisement feature vector and the client's portrait feature vector can be determined by using a similarity evaluation model based on machine learning.

[0063] In addition, a pre-trained retrieval model can be used to directly retrieve multiple target advertisement feature vectors that meet the requirements of similarity with the client's portrait feature vector from the vector database. The retrieval model can be a TF-IDF model, a DPR model, etc.

[0064] From the above description, it can be seen that by comparing the similarity between each advertisement feature vector in the vector database and the client's portrait feature vector, multiple target advertisement feature vectors whose similarity with the client's portrait feature vector meets the requirements are retrieved from the vector database.

[0065] After multiple target advertisement feature vectors whose similarity with the client's portrait feature vector meets the requirements are screened out, the advertisements to be placed corresponding to each of the multiple target advertisement feature vectors are determined according to the corresponding relationship between the advertisement feature vectors and the advertisements to be placed.

[0066] Next, according to the advertisements to be placed corresponding to the multiple target feature vectors and the client's portrait feature vectors, a target prompt text corresponding to the large language model is generated, and the target prompt text is input into the large language model to obtain a target advertisement that matches the client. The target prompt text is used to instruct the large language model to select a target advertisement from the multiple advertisements to be placed corresponding to the multiple target advertisement feature vectors.

[0067] In order to more quickly and accurately select the target advertisement that best matches the client from multiple advertisements to be delivered that meet the requirements for similarity with the client's portrait feature vector, a large language model can be used for processing.

[0068] In order to enable the large language model to better understand the task to be processed and select the target advertisement from the multiple advertisements to be placed corresponding to the multiple target advertisement feature vectors, before using the large language model to process the task, the target prompt text can be generated based on the advertisements to be placed and the client's portrait feature vectors corresponding to the multiple target feature vectors, and then the target prompt text is input into the large language model. Among them, the target prompt text can guide the large language model to understand the specific requirements, context environment and expected output format of the task. The target prompt text includes the defined task to be processed, such as "select the advertisement that best matches the specific client from a given list of advertisements". It can also include set context information, such as setting the context environment corresponding to the task to be processed for the large language model through the specific portrait information of the client and multiple advertisements to be placed, which helps the large language model understand the background of the task to be processed and make more accurate judgments. It can also include the format corresponding to the output result, so that the target prompt text can also guide the large language model on how to output the result, which helps to ensure that the content output by the model meets the task requirements.

[0069] From the above description, we can see that by carefully designing the prompt text, we can guide the model to accurately understand the task requirements and improve the efficiency and accuracy of task processing.

[0070] In summary, the embodiments of the present invention screen out multiple advertisements to be placed that match the client portrait by comparing the portrait feature vector with the advertisement feature vectors corresponding to multiple advertisements to be placed in the vector database, and generate a target prompt text based on the multiple advertisements to be placed and the portrait feature vector, and input the target prompt text into the large language model so that the large language model can better understand the semantic information, so as to screen out the target advertisement with the highest matching degree with the client from the multiple advertisements to be placed, that is, through two screenings, it is achieved to deliver advertisements to the client in a targeted manner according to the feature information of the client, which can not only improve the accuracy and conversion rate of advertisement delivery within the application, but also improve the user experience.

[0071] The above embodiment describes an implementation method for determining a client's portrait feature vector. In another optional embodiment, the portrait feature vector can be obtained by processing the user data using a user portrait model. Specifically, the user data corresponding to the client is obtained; the user data is input into a pre-trained user portrait model to obtain the user portrait information corresponding to the client; the user portrait information is vectorized to obtain the portrait feature vector corresponding to the user portrait information.

[0072] A large amount of user data may be collected. The user data includes basic user information, user behavior data, etc. The user data may be filtered, cleaned, etc. to obtain processed user data, and the processed user data may be used as training samples. Then, a suitable machine learning or deep learning model may be selected as a user portrait model. Next, the user data in the training sample is input into the user portrait model to obtain predicted user portrait information. The reference user portrait information corresponding to the user data is obtained, and the user portrait model is trained based on the reference user portrait information and the predicted user portrait information to obtain a trained user portrait model.

[0073] In order to more accurately determine the portrait feature vector corresponding to each client, in another optional embodiment, the user data can be processed by using a user classification model to obtain the user category, and the client's portrait feature vector can be further determined based on the user category and user data.

[0074] In order to understand the specific implementation process of determining the client's portrait feature vector based on the user classification model, you can combine the attached Figure 3 An exemplary description is given.

[0075] Figure 3 A flowchart of determining a client portrait feature vector provided by an embodiment of the present invention; based on the above embodiment, refer to the attached Figure 3 As shown, the method comprises the following steps:

[0076] 301. Obtain user data corresponding to the client.

[0077] 302. Input the user data into a pre-trained user classification model to obtain the user category corresponding to the client.

[0078] 303. Determine the user portrait information corresponding to the client according to the user category and user data.

[0079] 304. Perform vectorization processing on the user portrait information to obtain a portrait feature vector corresponding to the user portrait information.

[0080] When determining the portrait feature vector corresponding to the client, the user data corresponding to the client can be obtained first, and then the user data is input into a pre-trained user classification model to obtain the user category corresponding to the client.

[0081] The user classification model can learn the mapping relationship between different user data and multiple user categories. Optionally, the user classification model can determine the user category corresponding to the user data with the help of a softmax function. In specific implementation, the user classification model can be obtained by collecting training samples and learning the softmax function.

[0082] In an optional embodiment, the training process of the user classification model may first collect training samples of the target application, which include user data corresponding to multiple clients. Then, the reference user category corresponding to the user data in each sample is determined; the user data in each sample is input into the user classification model respectively, so as to train the user classification model according to the loss calculation result between the predicted user category output by the user classification model and the reference user category.

[0083] After obtaining the trained user classification model, the trained user classification model can be used to classify the user data corresponding to the client to determine the user category corresponding to the user data.

[0084] In addition, the training process of the user classification model adopts a continuous iteration mechanism. It continuously collects users' feedback on advertisements and the behavioral operations triggered by advertisements, and timely updates user portrait information based on users' real-time feedback and behavioral changes. Based on the timely updated user portrait information, the user classification model is continuously trained and learned, and the network parameters in the model are updated.

[0085] Next, the user portrait information corresponding to the client is determined based on the user category and user data. The user portrait information may include a detailed description of the user's basic information, interests, hobbies, purchasing behavior, usage habits, lifestyle, etc.

[0086] That is to say, when determining the user portrait information corresponding to the client, the user category corresponding to the client can be combined to more accurately define the user portrait information. Different categories of users have different behavior patterns, interest preferences and consumption needs. Combining user categories can refine user portrait information to make it closer to the user's actual situation.

[0087] Finally, the user portrait information is vectorized to obtain a portrait feature vector corresponding to the user portrait information. The vectorization process can convert complex user portrait information into a numerical form that can be processed by a computer, namely, a portrait feature vector.

[0088] In specific implementation, one-hot encoding (One-Hot Encoding), Bag of Words (BoW), TF-IDF (Term Frequency-Inverse Document Frequency), embedding technologies such as Word2Vec and BERT, etc. can be used to convert various attribute data in user data (such as basic information, behavioral data, text data, etc.) into user portrait feature vectors.

[0089] In an embodiment of the present invention, user data is classified by utilizing a user classification model to determine the corresponding user category, and user portrait information corresponding to the client is determined based on the user category and user data, and the user portrait information is vectorized to obtain a portrait feature vector. This ensures the accuracy of the determined user portrait information, and then the corresponding portrait feature vector can be determined based on the user portrait information. This allows the target advertisements that are more in line with user needs and interests to be screened out based on the client's portrait feature vector in the future.

[0090] After determining the portrait feature vector corresponding to the client, the portrait feature vector of the client can be stored in a preset storage area, so that the next time an advertisement is delivered to the client, the portrait feature vector of the client can be directly obtained from the preset storage area.

[0091] In addition, in actual applications, before performing target advertisement screening for the client, a vector database can be established, and the advertisement feature vectors corresponding to each determined advertisement to be placed can be stored in the vector database, so that when the client is detected to start the target application, multiple target advertisement feature vectors matching the client's portrait feature vector can be quickly retrieved from the vector database, and then multiple advertisements to be placed matching the client are retrieved. Then, the target advertisement that best matches the client is screened from the multiple advertisements to be placed.

[0092] Combined with Figure 4 The above specific implementation process is described in detail.

[0093] Figure 4 A flowchart of another advertisement delivery method provided by an embodiment of the present invention; based on the above embodiment, the method may further include the following steps:

[0094] 401. Get multiple advertisements to be delivered.

[0095] 402. Determine the advertisement type corresponding to each of the plurality of advertisements to be placed.

[0096] 403. Perform feature extraction on the multiple advertisements to be placed according to the advertisement types corresponding to the multiple advertisements to be placed, and determine advertisement feature vectors corresponding to the multiple advertisements to be placed.

[0097] 404. Store the advertisement feature vectors corresponding to the multiple advertisements to be placed in a vector database.

[0098] In specific implementation, multiple advertisements to be placed may be obtained from the advertisement data set. The advertisements to be placed may be text advertisements, image advertisements, video advertisements, etc. After obtaining the multiple advertisements to be placed, the advertisement types corresponding to the multiple advertisements to be placed are determined respectively. The advertisement types may include text advertisements, image advertisements, video advertisements, etc.

[0099] In addition, each category of advertisements can be further divided according to the corresponding advertisement content of each category of advertisements to determine the corresponding content type, so that the advertisement feature vector corresponding to each advertisement to be delivered can be determined more accurately, and then the target advertisement that best matches the client can be matched to the client more accurately, thereby improving the accuracy of advertisement delivery and user experience. Among them, the content type can include life advertisements, promotion advertisements, etc. For example, for advertisement A to be delivered, the advertisement A can be a generated advertisement in the image advertisement.

[0100] Next, according to the advertisement types corresponding to the multiple advertisements to be placed, feature extraction is performed on the multiple advertisements to be placed to determine the advertisement feature vectors corresponding to the multiple advertisements to be placed. In order to more accurately determine the advertisement feature vectors corresponding to each advertisement to be placed, in an embodiment of the present invention, different recognition models can be used for different categories of advertisements to be placed to determine the advertisement feature vectors corresponding to each advertisement.

[0101] Specifically, in an optional embodiment, according to the advertisement types corresponding to each of the multiple advertisements to be placed, feature extraction is performed on the multiple advertisements to be placed, and the implementation method of determining the advertisement feature vectors corresponding to each of the multiple advertisements to be placed can be: if the advertisement type corresponding to the target advertisement to be placed is a text advertisement, the target advertisement to be placed is input into a pre-trained text recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed. If the advertisement type corresponding to the target advertisement to be placed is an image advertisement, the target advertisement to be placed is input into a pre-trained image recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed. If the advertisement type corresponding to the target advertisement to be placed is a video advertisement, the target advertisement to be placed is input into a pre-trained video recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed.

[0102] The target advertisement to be delivered is any one of multiple advertisements to be delivered, and the processing procedures corresponding to each advertisement to be delivered are consistent. You can refer to the specific implementation process of the target advertisement delivery, which will not be described in detail.

[0103] In practice, a corresponding model can be built according to the type of advertisement, and the model can be trained to obtain a corresponding recognition model. Among them, the types of advertisements include at least text, image, video, etc., and text recognition models, image recognition models, and video recognition models can be created. In another optional embodiment, the types of advertisements can be further subdivided, and then corresponding recognition models can be created on this basis, such as life recognition models, promotion models, etc., which are not limited in the embodiments of the present invention.

[0104] In the specific implementation, a large number of training samples can be obtained first, including text advertisements, video advertisements, and image advertisements. These advertisements can come from, but are not limited to, television, the Internet, newspapers and magazines, etc. In order to ensure that the training data set is diverse and representative, it covers different types of advertisements and advertisement contents. Then, the training samples are classified to obtain advertisement samples of each category, and then, the corresponding recognition models are trained using the advertisement samples of the corresponding categories to obtain the corresponding trained recognition models.

[0105] Specifically, for the text recognition model, first obtain the advertising sample of the text class, perform text cleaning on the advertising sample, and remove irrelevant information such as HTML tags, special characters, stop words, etc. in the text. The cleaned text is divided into independent words or phrases for subsequent processing. Then, the text after word segmentation is processed to convert the text after word segmentation into a numerical vector. For example, a word embedding model can be used to convert the text after word segmentation into a numerical vector. Then, a text recognition model is constructed. Among them, the text recognition model can be an RNN model, a Transformer model, etc. And the text recognition model is learned and trained using the advertising samples.

[0106] In an optional embodiment, a long short-term memory network (LSTM) is selected as the RNN architecture for text processing. The RNN model includes an input layer, a hidden layer, and an output layer. The data vector converted from the text corresponding to the advertising sample can be input into the input layer of the RNN model as input data to obtain a vector table corresponding to each word or phrase. The input layer can encode the data vector into a vector form. Then, the vector representation corresponding to each word or phrase is input into the hidden layer to obtain a hidden state vector corresponding to the advertising sample. The hidden layer is responsible for processing the input sequence data. At each time step, the hidden layer receives the current input and the hidden state at the previous moment, and calculates the hidden state at the current moment. This hidden state can be regarded as the short-term memory of the RNN, which remembers past information and is used to influence subsequent decisions. Next, the hidden state vector corresponding to the advertising sample is input into the output layer to obtain the advertising feature vector corresponding to the advertising sample. The output layer is a fully connected layer that can convert the hidden state vector of the current time step into a feature vector of a fixed length.

[0107] It should be noted that the output layer can generate different outputs according to the advertising demand. For example, when the advertising demand corresponding to a certain advertisement to be placed is a lifestyle advertisement, a feature vector that better reflects the lifestyle characteristics can be generated.

[0108] In addition, in practical applications, the training samples can be divided into training sets, validation sets, and test sets. The appropriate loss function and optimizer can be selected according to the task requirements. The training set can be used to train the RNN model, and the model parameters can be updated through the back propagation algorithm. During the training process, the validation set can be used to monitor the performance of the model and adjust the parameters according to the requirements.

[0109] For the text recognition model, the acquired image-based advertising samples are first preprocessed to uniformly scale the advertising sample images of different sizes to a size that the model can handle, such as 200x200 pixels; the pixel values ​​of the images are normalized to the range of [0,1] or [-1,1] to accelerate the training process of the image recognition model. Training data is generated by randomly cropping, rotating, flipping, and other operations on the advertising sample images, which can increase the diversity of training data and improve the generalization ability of the image recognition model.

[0110] Then, an image recognition model is constructed. The image recognition model can be a CNN model. The CNN model includes convolutional layers, activation functions, and pooling layers. Construct a convolutional layer of the CNN model, and use multiple convolutional kernels to perform convolution operations on the image advertising samples to extract local features in the advertising sample images. Each convolutional kernel generates a feature map, which are combined to form the input of the next layer. Pooling layers (such as maximum pooling or average pooling) are used to reduce the dimension of feature maps while retaining key information, which helps reduce the computational complexity of the image recognition model and enhance its robustness to small changes in the input data. Between the convolutional layer and the pooling layer, an activation function (such as ReLU) is introduced to increase the nonlinear representation capability of the image recognition model.

[0111] The image recognition model is trained to obtain a trained image recognition model. During training, the CNN model can be trained using a labeled image training dataset, the model parameters can be updated through a back-propagation algorithm, and the performance of the model on the validation set can be monitored to select the appropriate number of training iterations and hyperparameters.

[0112] For the video recognition model, the obtained video advertising samples are first preprocessed. Preprocessing includes video format conversion, resolution adjustment, key frame extraction, etc. According to the requirements of the model, the video is converted into a format suitable for input. Then, a video recognition model is constructed. The video recognition model includes an image feature extraction module, a text feature extraction module, and a video understanding module. The image feature extraction module is used to extract visual features from video advertising samples, and use a pre-trained image feature extraction model (such as a CNN model) to extract visual features from the key frames of the video advertising samples. These visual features can include color, texture, shape, etc.

[0113] Next, the extracted visual features are combined according to the temporal relationship of the video to form a visual feature representation of the video sequence. The visual feature representation of the video sequence corresponding to the video advertisement sample is input into the text feature extraction module. The text feature extraction module uses text processing tools (such as word segmenters and word embedding models) to extract text features from the subtitles or descriptions in the video.

[0114] Then, the extracted text features are combined according to the temporal relationship of the video to form a text feature representation of the video sequence. The extracted visual features and text features are fused to form multimodal data. In addition, according to the requirements of the video understanding module, the fused multimodal data can be converted into an input format acceptable to the video understanding module, such as a specific data structure and dimension.

[0115] Finally, the multimodal data is input into the video understanding module to obtain the advertising feature vector corresponding to the advertising sample. The video understanding module can be a pre-trained Mini GPT4-Video model, which uses the Mini GPT4-Video model to calculate the multimodal data to extract key information and features from the advertising sample.

[0116] After the recognition models of each category are trained, the pre-trained recognition model can be used to perform feature extraction processing on the advertisements to be placed in the advertisement data set to obtain the advertisement feature vectors corresponding to each advertisement to be placed, and the advertisement feature vectors corresponding to each advertisement to be placed are stored in the vector database. For example, if the advertisements to be placed obtained are video advertisements, the target advertisements to be placed can be input into the pre-trained video recognition model to obtain the advertisement feature vectors corresponding to the target advertisements to be placed.

[0117] Specifically, in an optional embodiment, the target advertisement to be placed is subjected to image frame processing to obtain multiple frames of images; the multiple frames of images are respectively input into an image feature extraction module to obtain visual features corresponding to each of the multiple frames of images; the multiple frames of images are input into a text feature extraction module to perform text feature extraction on the subtitles in the images to obtain text features corresponding to each of the multiple frames of images; the visual features and text features are fused according to a temporal relationship to obtain fused features; the fused features are input into a video understanding module to extract key information from the target advertisement to be placed to obtain an advertisement feature vector corresponding to the target advertisement to be placed.

[0118] In the embodiment of the present invention, a large amount of advertising content is collected as a data set of advertisements to be placed, and the collected advertisements to be placed are classified into large categories to determine the advertisement types corresponding to the advertisements to be placed. Then, the large model-related technology is used to extract features of different types of advertisements to be placed, and the advertisement feature vectors corresponding to each advertisement to be placed are obtained. In this way, more accurate advertisement feature vectors corresponding to each advertisement to be placed can be obtained, thereby improving the matching results of the target advertisements. In addition, the determined advertisement feature vectors corresponding to each advertisement to be placed can also be stored in a vector database, so that the advertisement matching processing can be directly performed using each advertisement feature vector in the vector database in the future, which can improve the efficiency of advertisement placement.

[0119] After determining the client's profile feature vector, we can use the interests and hobbies in the client's user profile as query vectors based on RAG technology, and use similarity retrieval technology to find multiple ads that are most relevant to these interests and hobbies. To optimize this process, we can also generate prompt text after retrieval to guide the large language model to more accurately find the target ads that best match the user's interests.

[0120] Specifically, based on multiple advertisements to be placed related to the interests and hobbies of the client and the client's portrait feature vector, a target prompt text corresponding to the large language model is generated to guide the large language model to screen out the target advertisement that best matches the interests and hobbies of the client from multiple advertisements to be placed that match the interests and hobbies of the client. Figure 5 The specific implementation process of generating the target prompt text is illustrated by way of example.

[0121] Figure 5 A flowchart of generating a target prompt text corresponding to a large language model is provided in an embodiment of the present invention; based on the above embodiment, an embodiment of the present invention provides a specific implementation method of generating a target prompt text corresponding to a large language model according to a portrait feature vector of an advertisement to be placed and a client corresponding to multiple target advertisement vectors, and the method may include the following steps:

[0122] 501. Decode the portrait feature vector to obtain the user interest preference corresponding to the client.

[0123] 502. Obtain a preset prompt template, where the preset prompt template is used to define a content format for describing a plurality of advertisements to be placed and user interest preferences.

[0124] 503. Based on a preset prompt template, a target prompt text corresponding to the target feature vector and the user's interest preference is obtained, and the target prompt text is used to guide the large language model to perform reasoning.

[0125] When generating the target prompt text, the client's portrait feature vector can be decoded first to obtain the user interest preference corresponding to the client. A preset prompt template is obtained, and based on the preset prompt template, the target prompt text corresponding to the target feature vector and the user interest preference is obtained. The target prompt text is used to guide the large language model to perform reasoning.

[0126] The preset prompt template is used to define the content format for describing multiple advertisements to be placed and user interest preferences. In order to better generate the target prompt text, a corresponding preset prompt template can be designed first.

[0127] Specifically, the representation element corresponding to the preset prompt template can be determined. The representation element includes: an element for limiting the pending tasks corresponding to the large language model, an element for limiting the input information corresponding to the pending tasks, and an element for limiting the output information corresponding to the pending tasks. Among them, the pending tasks can be to select target advertisements that match the client from multiple advertisements to be placed, and based on the representation element corresponding to the preset prompt template, the pending tasks corresponding to the large language model are obtained.

[0128] For example, the task to be processed corresponding to the limited large language model is: selecting an advertisement that best matches the client from a given advertisement list.

[0129] In an embodiment of the present invention, by generating a target prompt text, the large language model can be better guided to select the target advertisement that best matches the client from multiple advertisements to be delivered, which can make the advertising delivery more targeted, improve the advertising delivery effect, and further improve the user experience.

[0130] Among them, the specific implementation process involved in the embodiments of the present invention can refer to the contents in the above embodiments, and will not be repeated here.

[0131] In addition, in actual applications, the client's feedback behavior on the target advertisement can be continuously collected, and the client's user portrait information and the client's portrait feature vector can be timely optimized and updated according to the user's real-time feedback and behavior changes.

[0132] The following will describe in detail the advertisement delivery device of one or more embodiments of the present invention. Those skilled in the art will appreciate that the advertisement delivery device can be configured using commercially available hardware components through the steps taught in this solution.

[0133] Figure 6 A schematic diagram of the structure of an advertisement delivery device provided by an embodiment of the present invention is shown in FIG. Figure 6 As shown, the device includes: a monitoring module 11, a retrieval module 12, a determination module 13, a generation module 14, and a processing module 15.

[0134] The monitoring module 11 is used to obtain a portrait feature vector of the client if it is detected that the client starts the target application.

[0135] The retrieval module 12 is used to retrieve a plurality of target advertisement feature vectors whose similarity with the portrait feature vector meets the requirements from a vector database, wherein the vector database stores advertisement feature vectors corresponding to a plurality of advertisements to be placed.

[0136] The determination module 13 is used to determine the advertisements to be placed corresponding to each of the plurality of target advertisement feature vectors according to the corresponding relationship between the advertisement feature vectors and the advertisements to be placed.

[0137] The generating module 14 is used to generate a target prompt text corresponding to the large language model according to the advertisement to be delivered and the portrait feature vector, wherein the target prompt text is used to instruct the large language model to filter out a target advertisement from a plurality of advertisements to be delivered corresponding to each of the plurality of target advertisement feature vectors.

[0138] The processing module 15 is used to input the target prompt text into a large language model to obtain a target advertisement that matches the client.

[0139] In an optional embodiment, the monitoring module 11 is specifically used to: obtain user data corresponding to the client; input the user data into a pre-trained user classification model to obtain a user category corresponding to the client; determine user portrait information corresponding to the client based on the user category and the user data; and vectorize the user portrait information to obtain a portrait feature vector corresponding to the user portrait information.

[0140] In an optional embodiment, the retrieval module 12 is specifically used to: respectively determine the similarity between the interest feature vector and each advertising feature vector in the vector database; sort the multiple advertising feature vectors in order of similarity from large to small, and select a preset number of target advertising feature vectors with the highest ranking.

[0141] In an optional embodiment, the retrieval module 12 is also used to: obtain multiple advertisements to be placed; determine the advertisement type corresponding to each of the multiple advertisements to be placed; extract features of the multiple advertisements to be placed according to the advertisement types corresponding to each of the multiple advertisements to be placed, and determine the advertisement feature vectors corresponding to each of the multiple advertisements to be placed; and store the advertisement feature vectors corresponding to each of the multiple advertisements to be placed in a vector database.

[0142] In an optional embodiment, the retrieval module 12 is further used for: if the advertisement type corresponding to the target advertisement to be placed is a text advertisement, the target advertisement to be placed is input into a pre-trained text recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed, and the target advertisement to be placed is any one of the multiple advertisements to be placed; if the advertisement type corresponding to the target advertisement to be placed is an image advertisement, the target advertisement to be placed is input into a pre-trained image recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed; if the advertisement type corresponding to the target advertisement to be placed is a video advertisement, the target advertisement to be placed is input into a pre-trained video recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed.

[0143] In an optional embodiment, the video recognition model includes an image feature extraction module, a text feature extraction module, and a video understanding module, and the retrieval module 12 is also used to: perform image frame processing on the target advertisement to be placed to obtain multiple frames of images; input the multiple frames of images into the image feature extraction module respectively to obtain visual features corresponding to each of the multiple frames of images; input the multiple frames of images into the text feature extraction module to extract text features of the subtitles in the image to obtain text features corresponding to each of the multiple frames of images; fuse the visual features and the text features according to a time sequence relationship to obtain fused features; input the fused features into the video understanding module to extract key information from the target advertisement to be placed and obtain an advertisement feature vector corresponding to the target advertisement to be placed.

[0144] In an optional embodiment, the generation module 14 is specifically used to: decode the portrait feature vector to obtain the user interest preference corresponding to the client; obtain a preset prompt template, the preset prompt template is used to limit the content format describing multiple advertisements to be delivered and the user interest preferences; based on the preset prompt template, obtain the target prompt text corresponding to the advertisements to be delivered and the user interest preferences, and the target prompt text is used to guide the large language model to perform reasoning.

[0145] In an optional embodiment, the generation module 14 is also used to: determine the representation element corresponding to the preset prompt template, the representation element includes: an element for limiting the pending tasks corresponding to the large language model, an element for limiting the input information corresponding to the pending tasks, and an element for limiting the output information corresponding to the pending tasks, the pending tasks being to filter out target advertisements that match the client from the multiple advertisements to be placed; based on the representation element, obtain the pending tasks corresponding to the large language model.

[0146] Figure 6 The device shown can execute the steps in the advertising delivery method in the aforementioned embodiment. The detailed execution process and technical effects can be found in the description of the aforementioned embodiment, which will not be repeated here.

[0147] The embodiment of the present invention also provides an electronic device, such as Figure 7 As shown, the electronic device may include: a processor 21, a memory 22, and a communication interface 23. The memory 22 stores executable codes, and when the executable codes are executed by the processor 21, the processor 21 implements the advertisement delivery method in the above-mentioned embodiment.

[0148] In addition, an embodiment of the present invention provides a non-transitory machine-readable storage medium, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor can at least implement the advertising delivery method provided in the aforementioned embodiment.

[0149] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. Those of ordinary skill in the art may understand and implement the present invention without creative effort.

[0150] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and software. Based on such an understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present invention can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An advertisement delivery method, characterized in that: include: If it is detected that the client starts the target application, a portrait feature vector of the client is obtained; Retrieving a plurality of target advertisement feature vectors whose similarity with the portrait feature vector meets the requirement from a vector database, wherein the vector database stores advertisement feature vectors corresponding to a plurality of advertisements to be placed; Determining the advertisements to be placed corresponding to each of the plurality of target advertisement feature vectors according to the corresponding relationship between the advertisement feature vectors and the advertisements to be placed; Generate a target prompt text corresponding to the large language model according to the advertisement to be placed and the portrait feature vector, wherein the target prompt text is used to instruct the large language model to filter out a target advertisement from a plurality of advertisements to be placed corresponding to each of the plurality of target advertisement feature vectors; The target prompt text is input into a large language model to obtain a target advertisement matching the client.

2. The method according to claim 1, characterized in that The obtaining of the client's portrait feature vector includes: Obtaining user data corresponding to the client; Inputting the user data into a pre-trained user classification model to obtain the user category corresponding to the client; Determine user portrait information corresponding to the client according to the user category and the user data; The user portrait information is vectorized to obtain a portrait feature vector corresponding to the user portrait information.

3. The method according to claim 1, characterized in that The portrait feature vector includes an interest feature vector, and retrieving from the vector database a plurality of target advertisement feature vectors whose similarity with the portrait feature vector meets the requirements includes: respectively determining the similarity between the interest feature vector and each advertisement feature vector in the vector database; The plurality of advertisement feature vectors are sorted in descending order of similarity, and a preset number of target advertisement feature vectors with top rankings are selected.

4. The method according to claim 1, characterized in that: Before retrieving from the vector database a plurality of target advertisement feature vectors whose similarity with the portrait feature vector meets the requirement, the method further comprises: Get multiple ads to be delivered; Determine the advertisement type corresponding to each of the plurality of advertisements to be placed; Extracting features of the multiple advertisements to be placed according to the advertisement types corresponding to the multiple advertisements to be placed, and determining advertisement feature vectors corresponding to the multiple advertisements to be placed; The advertisement feature vectors corresponding to each of the plurality of advertisements to be placed are stored in a vector database.

5. The method according to claim 4, characterized in that The extracting features of the multiple advertisements to be placed according to the advertisement types corresponding to the multiple advertisements to be placed, and determining the advertisement feature vectors corresponding to the multiple advertisements to be placed, comprises: If the advertisement type corresponding to the target advertisement to be placed is a text advertisement, the target advertisement to be placed is input into a pre-trained text recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed, and the target advertisement to be placed is any one of the multiple advertisements to be placed; If the advertisement type corresponding to the target advertisement to be placed is an image advertisement, inputting the target advertisement to be placed into a pre-trained image recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed; If the advertisement type corresponding to the target advertisement to be placed is a video advertisement, the target advertisement to be placed is input into a pre-trained video recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed.

6. The method according to claim 5, characterized in that The video recognition model includes an image feature extraction module, a text feature extraction module, and a video understanding module. The target advertisement to be placed is input into a pre-trained video recognition model to obtain an advertisement feature vector corresponding to the target advertisement to be placed, including: Performing image frame processing on the target advertisement to be placed to obtain multiple frames of images; Inputting the multiple frames of images into the image feature extraction module respectively to obtain visual features corresponding to the multiple frames of images; Inputting the multiple frames of images into the text feature extraction module, performing text feature extraction on the subtitles in the images, and obtaining text features corresponding to each of the multiple frames of images; According to the temporal relationship, the visual feature and the text feature are fused to obtain a fused feature; The fused features are input into the video understanding module to extract key information from the target advertisement to be delivered, and obtain an advertisement feature vector corresponding to the target advertisement to be delivered.

7. The method according to claim 1, characterized in that The step of generating a target prompt text corresponding to the large language model according to the advertisement to be placed and the portrait feature vector includes: Decoding the portrait feature vector to obtain the user interest preference corresponding to the client; Obtaining a preset prompt template, wherein the preset prompt template is used to define a content format for describing a plurality of advertisements to be placed and the user's interest preferences; Based on the preset prompt template, a target prompt text corresponding to the advertisement to be delivered and the user's interest preference is obtained, and the target prompt text is used to guide the large language model to perform reasoning.

8. The method according to claim 7, characterized in that The method further comprises: Determine a representation element corresponding to the preset prompt template, the representation element comprising: an element for defining a task to be processed corresponding to the large language model, an element for defining input information corresponding to the task to be processed, and an element for defining output information corresponding to the task to be processed, wherein the task to be processed is to select a target advertisement matching the client from the multiple advertisements to be placed; Based on the representation element, a task to be processed corresponding to the large language model is obtained.

9. An electronic device, characterized in that: include: A memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the advertising delivery method as described in any one of claims 1 to 8.

10. A non-transitory machine-readable storage medium, characterized in that: The non-transitory machine-readable storage medium stores executable code, and when the executable code is executed by a processor of an electronic device, the processor is caused to execute the advertisement delivery method according to any one of claims 1 to 8.

Images