Advertisement recommendation method based on artificial intelligence

Abstract

The invention provides an advertisement recommendation method based on artificial intelligence, and the method comprises the steps: screening information according to the advertisement preference of a user, and determining a candidate advertisement set, the information being an advertisement identifier or a historical interaction record inputted by the user; obtaining feature information of the candidate advertisements; and performing stream segmentation processing on a real-time advertisement stream of the advertisement putting platform based on the feature information to obtain an advertisement material set of each candidate advertisement. Wherein the stream segmentation processing comprises the steps of extracting local advertisement materials from the real-time advertisement stream, and generating an advertisement material set after duplicate removal in combination with cloud advertisement data returned by the recommendation server. According to the invention, personalized advertisement recommendation can be realized, and the user click rate and the advertisement putting effect are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and digital advertising technology, and more specifically, to an advertising recommendation method based on artificial intelligence. Background Art

[0002] In today's digital age, ad recommendation systems have become a crucial tool for internet platforms to attract user attention and improve advertising effectiveness. Existing ad recommendation systems primarily rely on users' browsing history, search keywords, and simple user profiles to deliver ads. These systems typically employ rule-based algorithms or simple machine learning models to analyze user behavioral data to predict user interests. For example, some systems recommend relevant ads based on the web content a user has visited or the types of products they have purchased over a period of time. These systems primarily collect user behavioral data, such as click-through rate and browsing time, and then use this data to train models to deliver personalized ad recommendations.

[0003] While existing ad recommendation systems can achieve personalized recommendations to a certain extent, they have some significant limitations. First, these systems often only make recommendations based on historical user behavior data and cannot capture changes in user interests in real time. Second, when processing ad creatives, existing ad recommendation systems typically only process single-modal data, such as text or images, and are unable to effectively integrate multimodal data, resulting in inaccurate recommendation results. Furthermore, existing ad recommendation systems also have shortcomings in the display and update of ad creatives, and are unable to dynamically adjust ad recommendations based on real-time user feedback.

[0004] In the process of implementing the embodiments of the present invention, there are at least the following problems or defects in the existing technology: the existing advertising recommendation system is unable to capture changes in user interests in real time, resulting in delayed recommendation results; it is unable to effectively integrate multimodal data, which limits the accuracy of recommendations; it lacks flexibility in the display and update of advertising materials and cannot be dynamically adjusted according to real-time user feedback. Summary of the Invention

[0005] The present invention provides an artificial intelligence-based advertising recommendation method, comprising: Determining a set of candidate advertisements for the user based on the user's advertisement preference screening information, wherein the set of candidate advertisements includes candidate advertisements of interest to the user; the advertisement preference screening information may be advertisement identification information input by the user on an advertisement selection interface, or may be a historical advertisement interaction record of the user; Obtaining feature information of each candidate advertisement in the candidate advertisement set; perform stream segmentation processing on the real-time advertisement stream of the advertisement delivery platform based on the feature information of each of the candidate advertisements, to obtain an advertisement material set of each of the candidate advertisements.

[0006] Further, the determining the candidate advertisement set of the user based on the advertisement preference screening information comprises: when the advertisement preference screening information is advertisement identification information input by the user on an advertisement selection interface, generating a candidate advertisement set based on each of the advertisement identifications in the advertisement identification information; when the advertisement preference screening information is a historical advertisement interaction record of the user, obtaining a preferred advertisement sequence of the user from a historical advertisement library based on the historical advertisement interaction record; identifying each advertisement element in the preferred advertisement sequence, screening a candidate advertisement in each of the advertisement elements, and generating a candidate advertisement set based on the screened candidate advertisements.

[0007] Further, the screening a candidate advertisement in each of the advertisement elements comprises: for each of the advertisement elements, determining a weight value of the advertisement element based on the historical advertisement interaction record and the preferred advertisement sequence, and determining an estimated interest value of the user for the advertisement element based on the weight value of the advertisement element; selecting a candidate advertisement from each of the advertisement elements based on the estimated interest value of each of the advertisement elements.

[0008] Further, the performing stream segmentation processing on the real-time advertisement stream of the advertisement delivery platform based on the feature information of each of the candidate advertisements, to obtain an advertisement material set of each of the candidate advertisements, comprises: for each of the candidate advertisements, extracting each first advertisement material containing the candidate advertisement from the real-time advertisement stream based on the feature information of the candidate advertisement; uploading the feature information of each of the candidate advertisements and the real-time advertisement stream of the advertisement delivery platform to a recommendation server to obtain cloud advertisement data of each of the candidate advertisements fed back by the recommendation server, the cloud advertisement data of each of the candidate advertisements containing each second advertisement material corresponding to the feature information of the candidate advertisement extracted from the real-time advertisement stream by the recommendation server; for each of the candidate advertisements, performing deduplication processing on each first advertisement material and each second advertisement material of the candidate advertisement to obtain each advertisement material after deduplication, and generating an advertisement material set of the candidate advertisement based on each advertisement material after deduplication.

[0009] Further, the generating an advertisement material set of the candidate advertisement based on each advertisement material after deduplication comprises: determine each of the de-duplicated advertisement material as a target advertisement material; group the target advertisement materials based on preset advertisement grouping rules to obtain a plurality of advertisement material groups, wherein each of the advertisement material groups comprises at least one target advertisement material.

[0010] Further, the method further comprises: displaying the advertisement material sets of the candidate advertisements to the user; determining a target advertisement material set from the advertisement material sets based on a trigger command received from the user; determining a target advertisement material group from the target advertisement material set based on the trigger command; outputting each of the advertisement materials in the target advertisement material group to the user.

[0011] Further, the de-duplication of the first advertisement materials and the second advertisement materials of the candidate advertisements comprises: extracting a first feature vector of each of the first advertisement materials and a second feature vector of each of the second advertisement materials; calculating a similarity value between each of the first feature vectors and all of the second feature vectors; when there is a feature vector pair with a similarity value exceeding a preset threshold, deleting the advertisement material with a lower priority in the feature vector pair.

[0012] Further, the method further comprises: parsing multi-modal data of the candidate advertisements, wherein the multi-modal data comprises text description data, image pixel data, and video frame sequence data; generating a text feature vector, an image feature vector, and a video feature vector by using a pre-trained feature extraction model; fusing the text feature vector, the image feature vector, and the video feature vector to generate an advertisement feature set.

[0013] Further, the fusion process adopts an attention weighting mechanism, comprising: dynamically adjusting fusion weight coefficients of the text feature vector, the image feature vector, and the video feature vector according to a user device type; when the user device is a mobile terminal, setting the weight coefficient of the image feature vector to 1.2 times of a reference weight.

[0014] Further, the method further comprises: monitoring, in real time, a user's dwell time data and click behavior data on the advertisement materials; When it is detected that the staying duration of the continuous N advertisement materials is lower than the preset duration threshold and there is no click behavior, a regeneration process of the candidate advertisement set is triggered; the regeneration process includes emptying the current candidate advertisement set and returning to performing the operation of determining the candidate advertisement set based on the user's advertisement preference screening information.

[0015] The above-mentioned embodiments of the present application have at least the following beneficial effects: 1. By combining the user's active input advertisement identification information with the historical interaction record to generate the candidate advertisement set, the problem of incomplete interest coverage caused by the dependence of the traditional recommendation system on single behavior data is solved, both the user's explicit preference demand and the potential interest are mined, and the relevance of the candidate advertisement is improved. 2. The local and cloud collaborative flow segmentation processing mechanism is adopted to quickly match the advertisement materials based on multi-dimensional feature information in the real-time advertisement flow, the problem of low advertisement retrieval efficiency in the high concurrency scene is solved, and the allocation accuracy of the advertisement resources is optimized by avoiding repeated content display through the deduplication processing. 3. The multi-modal feature fusion and dynamic weight adjustment technology is introduced, the fusion strategy of text, image and video features is adaptively optimized for different terminal devices, the problem of poor adaptability of cross-media advertisement materials is solved, the recommendation result is more in line with the user's device characteristics and interaction habits, and the advertisement presentation effect is improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when read in conjunction with the accompanying drawings. In the drawings, several embodiments of the present application are illustrated by way of example in which: Figure 1 A flowchart of an advertisement recommendation method based on artificial intelligence provided by an embodiment of the present application. DETAILED DESCRIPTION

[0017] The principles and spirits of the present application will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the present application, and do not limit the scope of the present application in any way. On the contrary, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0018] Those skilled in the art know that the embodiments of the present application can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present application can be embodied in the following forms: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0019] It should be noted that the number of any elements in the drawings is used for illustration and not limitation, and any naming is only used for differentiation and does not have any limiting meaning.

[0020] Reference will be made to the drawings Figure 1 , Figure 1 The flowchart of the AI-based advertisement recommendation method provided by an embodiment of the present application is shown in FIG. 1. As shown in FIG. 1, an AI-based advertisement recommendation method includes the following steps: Figure 1 S1, filtering information based on the user's advertisement preference to determine the user's candidate advertisement set, which contains each candidate advertisement that the user is interested in; the advertisement preference filtering information is the advertisement identifier information input by the user on the advertisement selection interface, or the user's historical advertisement interaction record; S2, obtaining the feature information of each candidate advertisement in the candidate advertisement set; S3, based on the feature information of each candidate advertisement, performing stream segmentation processing on the real-time advertisement stream of the advertisement delivery platform to obtain the advertisement material set of each candidate advertisement.

[0021] It should be noted that the present application proposes an AI-based advertisement recommendation method. The core of this method is to determine the candidate advertisement set through the user's advertisement preference filtering information. The advertisement preference filtering information here can be the advertisement identifier information actively input by the user on the advertisement selection interface, or the preference data automatically filtered by the system according to the user's historical advertisement interaction record. The so-called advertisement identifier information refers to the identifier of the advertisement type or specific advertisement selected by the user on the advertisement selection interface, such as the ID or keyword of the advertisement. The historical advertisement interaction record covers the user's past interaction behaviors with the advertisement, such as clicking, browsing time, collecting, etc. After determining the candidate advertisement set, the system will obtain the feature information of each candidate advertisement, which refers to various data that can describe the content and attributes of the advertisement, such as the theme, style, target audience, etc. of the advertisement. Then, based on these feature information, the real-time advertisement stream of the advertisement delivery platform is processed by stream segmentation to obtain the advertisement material set of each candidate advertisement. The so-called stream segmentation processing refers to dividing the continuous advertisement stream into multiple independent advertisement material segments according to certain rules, so as to facilitate subsequent personalized recommendation.

[0022] ​Specifically, the candidate advertisement set refers to a group of advertisements that are preliminarily screened according to the advertisement preference screening information of the user and are likely to meet the interests of the user. If the advertisement preference screening information of the user is the advertisement identification information, the system will directly generate the candidate advertisement set according to the identification information, that is, the advertisements that the user explicitly indicates interest in are included in the set. If the advertisement preference screening information of the user is the historical advertisement interaction record, the system will extract the advertisement sequence that the user has interacted with from the historical advertisement library, further identify each advertisement element in the advertisement sequence, such as the theme and style of the advertisement, and screen the candidate advertisements based on these elements to generate the candidate advertisement set. When obtaining the feature information of the candidate advertisement, the multi-modal data of the candidate advertisement needs to be parsed, including text description data, image pixel data, and video frame sequence data, etc. The pre-trained feature extraction model is used to generate text feature vectors, image feature vectors, and video feature vectors respectively, and finally the feature vectors are fused to generate the advertisement feature set. When performing stream segmentation processing on the real-time advertisement stream, for each candidate advertisement, based on its feature information, each first advertisement material containing the candidate advertisement is extracted from the real-time advertisement stream, and the feature information of the candidate advertisement and the real-time advertisement stream are uploaded to the recommendation server to obtain the cloud advertisement data of each candidate advertisement fed back by the recommendation server, which contains each second advertisement material corresponding to the feature information of the candidate advertisement extracted from the real-time advertisement stream. Then, the first advertisement material and the second advertisement material of each candidate advertisement are de-duplicated to obtain the de-duplicated advertisement material, and the advertisement material set of the candidate advertisement is generated based on these materials.

[0023] Preferably, a weight value can be set for each advertising element when determining the estimated interest value of the user to the advertising element, which can be determined according to the historical advertising interaction records and the preferred advertising sequence of the user, for example, if the user often clicks on a certain type of advertising, the weight value of this type of advertising element will be relatively high. The calculation of the estimated interest value can be based on the weight value and the historical interaction frequency of the user to similar advertising elements, etc. In extracting the feature vector of the advertising material, the feature extraction model can use the convolutional neural network CNN in deep learning to process image and video data, and the recurrent neural network RNN or its variant long short-term memory network LSTM to process text data. In the fusion of the text feature vector, the image feature vector and the video feature vector to generate the advertising feature set, an attention weighting mechanism can be used to dynamically adjust the fusion weight coefficient of each feature vector according to the user device type, for example, when the user device is a mobile terminal, considering that mobile device users tend to quickly obtain information through images, the weight coefficient of the image feature vector can be set to 1.2 times the baseline weight. In the deduplication processing of the advertising material, the feature vector of each advertising material can be extracted, the similarity value between the feature vectors is calculated, and when the similarity value exceeds a preset threshold, the advertising material with lower priority is deleted, and the priority can be determined according to the novelty, relevance and other factors of the advertising material.

[0024] In some embodiments, the user-based advertising preference screening information is used to determine the candidate advertising set of the user, including: When the advertising preference screening information is the advertising identifier information input by the user in the advertising selection interface, a candidate advertising set is generated based on each advertising identifier in the advertising identifier information; When the advertising preference screening information is the historical advertising interaction record of the user, a preferred advertising sequence of the user is obtained from a historical advertising library based on the historical advertising interaction record; Each advertising element in the preferred advertising sequence is identified, candidate advertisements are screened in each advertising element, and a candidate advertising set is generated based on the screened candidate advertisements.

[0025] It is necessary to explain that the present application adopts different strategies in determining the user candidate advertisement set according to the user-provided advertisement preference screening information. When the user actively inputs advertisement identification information in the advertisement selection interface, the system will directly generate the candidate advertisement set according to these identifiers. The advertisement identification information here refers to the user's explicit expression of interest in specific advertisements, for example, the user inputs the advertisement categories, brand names or identifiers of specific advertisement activities that they are interested in. On the other hand, when the user's advertisement preference screening information is based on their historical advertisement interaction records, the system will extract the user's preferred advertisement sequence from the historical advertisement library. The so-called historical advertisement interaction record refers to the user's past interaction behavior with advertisements, such as clicking, browsing, collecting, etc., and the preferred advertisement sequence is a series of advertisements that the user may be interested in, which is analyzed according to these interaction behaviors. The system will identify each advertisement element in these sequences and screen out candidate advertisements from them to finally generate the candidate advertisement set. This process involves analysis and processing of user behavior data, aiming to more accurately capture the user's interest preferences and provide personalized advertisement recommendations.

[0026] Specifically, when the user's advertisement preference screening information is advertisement identification information, the system will take each advertisement identifier input by the user as the screening basis and directly match the corresponding advertisements from the advertisement library to generate the candidate advertisement set. The advertisement identifier here can be the ID, keyword, brand name, etc. of the advertisement, and the system will accurately locate the advertisements that the user is interested in according to these identifiers. When the user's advertisement preference screening information is historical advertisement interaction records, the system will extract the user's preferred advertisement sequence from the historical advertisement library. This sequence is generated according to the user's past behavior data and reflects the user's historical interests. The system will identify each advertisement element in the sequence, such as the theme, style, target audience, etc. of the advertisement, and screen out candidate advertisements based on these elements. The screening process will consider the weight values of the advertisement elements, which are determined according to the user's historical interaction frequency and intensity with different advertisement elements, for example, if the user frequently clicks on a certain type of advertisement, then the weight value of this type of advertisement element will be relatively high. In this way, the system can more accurately screen out candidate advertisements that meet the user's interests and generate the candidate advertisement set.

[0027] Preferably, in screening the candidate advertisements, the system can further refine the screening process. For example, for each advertisement element, the system can determine a weight value of the advertisement element based on the user's historical advertisement interaction records and preferred advertisement sequences. The determination of the weight value can be achieved by analyzing the user's behavior data such as click rate, browsing time, number of collections, etc. on different advertisement elements. Specifically, a base weight can be set for each advertisement element, and the base weight can be adjusted according to the user's behavior data. For example, if the user has a high click rate on a certain advertisement element, the weight value of the element can be appropriately increased. In determining the user's estimated interest value for the advertisement element, factors such as the weight value of the advertisement element and the user's historical interaction frequency on similar advertisement elements can be combined for calculation. The calculation of the estimated interest value can use a weighted average method, in which the contribution of the advertisement element with a higher weight value to the estimated interest value is greater. In this way, the system can more accurately assess the user's interest in each advertisement element and select candidate advertisements accordingly. In addition, in generating the candidate advertisement set, the system can also consider the novelty and diversity of the advertisements to avoid recommending too similar advertisements, thereby improving the richness of the recommendations and user experience.

[0028] In some embodiments, the screening of the candidate advertisements in each of the advertisement elements comprises: For each of the advertisement elements, determining a weight value of the advertisement element based on the historical advertisement interaction records and the preferred advertisement sequences, and determining an estimated interest value of the user for the advertisement element based on the weight value of the advertisement element; Selecting the candidate advertisements from each of the advertisement elements based on the estimated interest values of each of the advertisement elements.

[0029] It should be noted that in screening the candidate advertisements, the present application determines the weight value of each advertisement element according to the user's historical advertisement interaction records and preferred advertisement sequences, and calculates the user's estimated interest value for each advertisement element based on these weight values. The advertisement element here refers to the basic features that constitute the advertisement, such as the theme, style, target audience, etc. of the advertisement. By analyzing the user's historical interaction behavior on different advertisement elements, the system can assess the user's interest in these elements and select candidate advertisements accordingly. This method can more accurately reflect the user's interest preferences, thereby improving the relevance and personalization of advertisement recommendations.

[0030] Specifically, when the user's advertisement preference screening information is historical advertisement interaction records, the system extracts the user's preferred advertisement sequence from the historical advertisement library. This sequence is generated based on the user's past behavior data and reflects the user's historical interests. The system identifies each advertisement element in the sequence and determines a weight value for each advertisement element. The determination of the weight value is based on the user's historical interaction frequency and intensity with different advertisement elements, for example, if the user frequently clicks on a certain type of advertisement, the weight value of this type of advertisement element will be relatively high. The calculation of the estimated interest value can use the weighted average method, in which the advertisement elements with higher weight values contribute more to the estimated interest value. Specific parameter settings can include click-through rate, browsing time, number of collections, etc., which reflect the user's interest in the advertisement elements. For example, the click-through rate can represent the user's interest intensity in the advertisement, the browsing time can represent the user's attention to the content of the advertisement, and the number of collections can represent the user's long-term interest in the advertisement. Through the comprehensive analysis of these parameters, the system can more accurately evaluate the user's interest in each advertisement element.

[0031] Preferably, in the screening of candidate advertisements, the system can further refine the screening process. For example, for each advertisement element, the system can determine the weight value of the advertisement element based on the user's historical advertisement interaction records and preferred advertisement sequence. The determination of the weight value can be achieved by analyzing the user's behavior data such as click-through rate, browsing time, number of collections, etc. Specifically, a base weight can be set for each advertisement element, and the base weight can be adjusted according to the user's behavior data. For example, if the user has a high click-through rate on a certain advertisement element, the weight value of that element can be appropriately increased. In determining the user's estimated interest value for the advertisement element, factors such as the weight value of the advertisement element and the user's historical interaction frequency with similar advertisement elements can be combined to calculate the estimated interest value. The calculation of the estimated interest value can use the weighted average method, in which the advertisement elements with higher weight values contribute more to the estimated interest value.

[0032] Further, the system can also introduce a time decay factor to consider the change of user interest over time. For example, for older interaction behaviors, their weights can be appropriately reduced, while for newer interaction behaviors, their weights can be appropriately increased. In this way, the system can more accurately evaluate the user's current interest in each advertisement element and select candidate advertisements accordingly.

[0033] In some embodiments, the stream segmentation processing of the real-time advertisement stream based on the feature information of each candidate advertisement to obtain the advertisement material set of each candidate advertisement comprises: For each candidate advertisement, based on the feature information of the candidate advertisement, each first advertisement material containing the candidate advertisement is extracted from the real-time advertisement stream. characteristic information of each of the candidate advertisements and the real-time advertisement stream of the advertisement delivery platform to a recommendation server to obtain cloud advertisement data of each of the candidate advertisements fed back by the recommendation server, the cloud advertisement data of each of the candidate advertisements containing each second advertisement material corresponding to the characteristic information of the candidate advertisement extracted by the recommendation server from the real-time advertisement stream; For each of the candidate advertisements, the de-duplication processing is performed on each first advertisement material and each second advertisement material of the candidate advertisement to obtain each de-duplicated advertisement material, and an advertisement material set of the candidate advertisement is generated based on each de-duplicated advertisement material.

[0034] It should be noted that when generating the advertisement material set of the candidate advertisement, the real-time advertisement stream of the advertisement delivery platform is subjected to stream segmentation processing based on the characteristic information of each candidate advertisement. The characteristic information refers to various data capable of describing the content and attributes of the advertisement, such as the theme, style, and target audience of the advertisement. The real-time advertisement stream refers to the sequence of advertisements currently being played or about to be played by the advertisement delivery platform. Through the stream segmentation processing, the system can extract the advertisement material matching the characteristic information of the candidate advertisement from the real-time advertisement stream, thereby generating an advertisement material set containing multiple advertisement materials for each candidate advertisement. The purpose of this process is to ensure that the advertisement material recommended to the user is diverse and relevant, while avoiding the repeated recommendation of the same advertisement material.

[0035] Specifically, for each candidate advertisement, the system extracts each first advertisement material containing the candidate advertisement from the real-time advertisement stream based on the characteristic information thereof. The first advertisement material refers to the advertisement material matching the characteristic information of the candidate advertisement directly extracted by the system from the real-time advertisement stream. At the same time, the system uploads the characteristic information of each candidate advertisement and the real-time advertisement stream to the recommendation server to obtain the cloud advertisement data of each candidate advertisement fed back by the recommendation server. The cloud advertisement data contains each second advertisement material corresponding to the characteristic information of the candidate advertisement extracted by the recommendation server from the real-time advertisement stream. The second advertisement material refers to the advertisement material extracted by the recommendation server from the cloud database according to the characteristic information of the candidate advertisement. For each candidate advertisement, the system performs de-duplication processing on the first advertisement material and the second advertisement material to avoid the repeated recommendation of the same advertisement material. The de-duplication processing can be realized by comparing the unique identifier or content feature vector of the advertisement material, ensuring that the advertisement material set of each candidate advertisement does not contain duplicate advertisement materials.

[0036] Preferably, when extracting the advertising material, the system can adopt a multi-modal data fusion technology to improve the accuracy of extraction. For example, for each candidate advertisement, the system can parse its multi-modal data, including text description data, image pixel data, and video frame sequence data. Through a pre-trained feature extraction model, text feature vectors, image feature vectors, and video feature vectors are generated respectively, and these feature vectors are fused to generate an advertising feature set. In the deduplication process, the system can extract the first feature vector of each first advertising material and the second feature vector of each second advertising material, and calculate the similarity value of each first feature vector and all second feature vectors. When the similarity value exceeds a preset threshold, the system can delete the advertising material with lower priority in the feature vector pair. The priority can be determined according to factors such as novelty and relevance of the advertising material. For example, if an advertising material is recently updated and highly relevant to user interest, its priority will be higher. In this way, the system can more accurately generate the advertising material set of each candidate advertisement, thereby improving the quality of advertising recommendation and user experience.

[0037] In some embodiments, the generating of the advertising material set of the candidate advertisement based on the deduplicated advertising materials comprises: determining each deduplicated advertising material as a target advertising material; grouping each target advertising material based on a preset advertising grouping rule to obtain a plurality of advertising material groups, each advertising material group containing at least one target advertising material.

[0038] It should be noted that when generating the advertising material set of the candidate advertisement, the system determines each deduplicated advertising material as a target advertising material, and groups these target advertising materials based on a preset advertising grouping rule to obtain a plurality of advertising material groups. The target advertising material here refers to an advertising material that is confirmed to be non-repetitive and meets the characteristics of the candidate advertisement after deduplication. The advertising grouping rule refers to a set of rules preset by the system for classifying and organizing advertising materials according to certain logic, so as to more effectively display them to users. In this way, the system can better manage advertising materials and improve the diversity of recommendations and user experience.

[0039] Specifically, when the system completes the deduplication of the first and second advertisement materials of the candidate advertisements, each independent advertisement material obtained is determined as a target advertisement material. These target advertisement materials will be grouped according to preset advertisement grouping rules. The advertisement grouping rules can be based on various factors, such as the theme, style, target audience, and type of the advertisement. For example, if the theme of the advertisement is tourism, all advertisement materials related to tourism can be grouped into one group; if the style of the advertisement is promotion, all advertisement materials of the promotion type can be grouped into another group. The specific parameter settings of the grouping rules can be determined according to the needs of the advertisement delivery platform and the behavior data of the user. For example, the grouping rules can be set according to the user's historical browsing records. If the user frequently browses technology-related advertisements, the system can group technology-related advertisement materials into a group and preferentially display them to the user.

[0040] Preferably, during the grouping process, the system can further refine the grouping rules to improve the accuracy and diversity of recommendations. For example, the system can dynamically adjust the grouping rules based on factors such as the novelty, relevance, and user feedback of the advertisement materials. For advertisement materials with high novelty, the system can group them separately and preferentially recommend them to the user. For advertisement materials with high relevance, the system can group them into the same group so that the user can browse multiple related advertisements at once.

[0041] Further, the system can also introduce a user feedback mechanism to adjust the grouping rules based on behavior data such as click-through rate and browsing duration. For example, if the click-through rate of a certain group of advertisement materials is low, the system can adjust the grouping rules to redistribute the advertisement materials in that group to other groups or optimize that group by adding new advertisement materials. In this way, the system can more flexibly manage advertisement materials and improve the relevance of recommendations and user experience.

[0042] In some embodiments, further comprising: displaying the set of advertisement materials of each of the candidate advertisements to the user; when receiving a trigger command of the user feedback, determining a set of target advertisement materials from each of the set of advertisement materials based on the trigger command; determining a group of target advertisement materials from the set of target advertisement materials using the trigger command; outputting each of the advertisement materials in the group of target advertisement materials to the user.

[0043] It should be noted that after generating the candidate advertisement material set, the application will display these sets to the user, and determine the target advertisement material set and the target advertisement material group according to the feedback trigger command of the user. The trigger command here refers to the interactive behavior of the user when browsing the advertisement material set, such as clicking, sliding, staying, etc. Through the trigger command of the user, the system can determine the advertisement material set and the specific advertisement material group that the user is interested in, so as to more accurately output the advertisement material that the user is interested in. The purpose of this process is to improve the interactivity of the user and the advertisement and the relevance of the advertisement recommendation.

[0044] Specifically, after the system generates the candidate advertisement material set, the system will display these sets to the user. The user will perform various interactive behaviors during the browsing process, which are recognized by the system as trigger commands. For example, the user clicks on a certain advertisement material set, or stays on a certain advertisement material for a long time, which are regarded as the user's interest in the advertisement material. The system will determine the target advertisement material set in each advertisement material set according to these trigger commands. Further, the system will use the trigger command to determine the target advertisement material group in the target advertisement material set. The target advertisement material set here refers to the advertisement material set selected by the user through the trigger command, and the target advertisement material group refers to the advertisement material group further selected by the user in the target advertisement material set. The system will output each advertisement material in the target advertisement material group to the user to meet the user's interest needs.

[0045] Preferably, when processing the trigger command of the user, the system can further refine the interactive logic. For example, the system can set different trigger thresholds, such as the number of clicks, the length of stay, etc., to judge the degree of interest of the user. If the user stays on a certain advertisement material for more than a certain length of time, or clicks on a certain advertisement material set multiple times, the system can regard these behaviors as high interest performance and preferentially recommend related advertisement materials.

[0046] Further, the system can introduce a machine learning model to analyze the trigger command of the user. The model can predict the interest preferences of the user according to the historical behavior data and real-time interaction data of the user. For example, the model can input the user's click history, browsing time, stay position, etc. Parameters, learn the user's interest mode through training, so as to more accurately determine the target advertisement material set and the target advertisement material group. When outputting the advertisement material, the system can also optimize according to the device type and network status of the user, for example, preferentially outputting picture and short video materials on mobile devices to improve loading speed and user experience.

[0047] In some embodiments, the de-duplication processing of the candidate advertisement material set includes: extract a first feature vector of each first advertising material and a second feature vector of each second advertising material; calculate a similarity value between each first feature vector and all second feature vectors; when there is a pair of feature vectors with a similarity value exceeding a preset threshold, delete the advertising material with lower priority in the pair.

[0048] It should be noted that when the system performs deduplication on each first advertising material and each second advertising material of the candidate advertisements, it extracts the feature vectors of each advertising material and calculates the similarity values between these feature vectors. When there is a pair of feature vectors with a similarity value exceeding a preset threshold, the system deletes the advertising material with lower priority in the pair. The feature vector here refers to a vector obtained through a feature extraction model, which can represent the content and attributes of the advertising material, and the similarity value refers to a numerical value that measures the similarity between two feature vectors. In this way, the system can identify and remove duplicate or highly similar advertising materials, thereby improving the diversity of advertising recommendations and user experience.

[0049] Specifically, when the system performs deduplication, it first extracts a first feature vector of each first advertising material and a second feature vector of each second advertising material. The extraction of feature vectors can be achieved through pre-trained deep learning models, which can process data in multiple modalities such as text, images, and videos. For example, for image data, a convolutional neural network (CNN) can be used to extract image feature vectors; for text data, a recurrent neural network (RNN) or its variant long short-term memory network (LSTM) can be used to extract text feature vectors. After extracting the feature vectors, the system calculates the similarity values between each first feature vector and all second feature vectors. The calculation of similarity values can use methods such as cosine similarity, which measures the similarity between two vectors by calculating the angle between them. When the similarity value exceeds the preset threshold, the system identifies that the advertising materials corresponding to this pair of feature vectors are duplicate or highly similar. At this time, the system deletes the advertising material with lower priority according to the priority rules. The priority rules can be set based on factors such as novelty, relevance, user feedback, etc. of the advertising material, for example, newly released advertising materials usually have higher priority.

[0050] Preferably, during the deduplication process, the system can further optimize the extraction of feature vectors and the calculation of similarity values. For example, in order to improve the representational ability of feature vectors, the system can adopt a multimodal fusion method to fuse data from multiple modalities such as text, images, and videos into one feature vector. When constructing a feature extraction model, data such as the text description, image content, and video frame sequence of the advertising material can be input, and through model training and learning, a feature vector that can comprehensively represent the content of the advertising material can be obtained. When calculating the similarity value, in addition to using cosine similarity, other similarity measurement methods, such as Euclidean distance, can also be combined to more comprehensively evaluate the similarity between advertising materials.

[0051] Furthermore, the system can dynamically adjust the similarity threshold to optimize deduplication based on the number and diversity of creatives. For example, when there are a large number of creatives, the similarity threshold can be appropriately increased to more strictly remove duplicates; when there are fewer creatives, the similarity threshold can be appropriately lowered to avoid insufficient creatives due to excessive deduplication. Through these optimization measures, the system can more effectively remove duplicate or highly similar creatives, improving the quality of ad recommendations and user experience.

[0052] In some embodiments, obtaining characteristic information of candidate advertisements includes: Parsing multimodal data of the candidate advertisement, the multimodal data including text description data, image pixel data, and video frame sequence data; Generate text feature vectors, image feature vectors, and video feature vectors respectively through the pre-trained feature extraction model; The text feature vector, image feature vector and video feature vector are integrated to generate an advertisement feature set.

[0053] It should be noted that when the present invention obtains the feature information of the candidate advertisement, it will parse the multimodal data of the candidate advertisement, including text description data, image pixel data and video frame sequence data. The pre-trained feature extraction model generates text feature vectors, image feature vectors and video feature vectors respectively, and these feature vectors are fused to generate an advertisement feature set. The multimodal data here refers to data containing a variety of different types of information, such as text, images and videos, which can describe the content and attributes of the advertisement from different angles. The feature extraction model refers to a model trained by machine learning or deep learning methods, which can convert multimodal data into feature vectors for subsequent processing and analysis. By fusing feature vectors of different modalities, the system can more comprehensively characterize the characteristics of the advertisement, thereby improving the accuracy and relevance of advertisement recommendations.

[0054] Specifically, analyzing the multi-modal data of the candidate advertisement includes extracting keywords and semantic information from the text description data, analyzing visual features such as color, texture, and shape from the image pixel data, and processing dynamic information and scene changes from the video frame sequence data. Text feature vectors can be generated through natural language processing techniques, such as using word embedding models like Word2Vec to convert words in the text into vectors, and then extracting semantic features of the text through text classification or clustering algorithms. Image feature vectors can be extracted using a convolutional neural network (CNN), which can automatically learn local features and global structures in images. Video feature vectors can be obtained by sampling video frame sequences and then using CNN or other video processing models to extract features from each frame, and then aggregating these features to obtain the video feature vector. When fusing these feature vectors, a simple concatenation method can be used, or a more complex fusion strategy such as weighted summation or attention mechanism can be used to highlight the importance of different modal data. The fused advertisement feature set can more comprehensively reflect the multi-aspect features of the advertisement, providing more rich information for subsequent advertisement recommendation.

[0055] Preferably, when constructing the feature extraction model, a deep learning framework such as TensorFlow or PyTorch can be used to train the model. Input parameters can include text sequences of text description data, image matrices of image pixel data, and video frame sequences of video frame sequence data. During model training, labeled advertisement data sets can be used to supervise model learning to ensure that the model can accurately extract features of different modal data. For example, for text data, a labeled text classification data set can be used to train the text feature extraction model; for image and video data, labeled image recognition and video classification data sets can be used to train the corresponding feature extraction model. When fusing feature vectors, an attention mechanism can be introduced to dynamically adjust the weights of different modal features according to user interest preferences. For example, if the user pays more attention to the visual effects of the advertisement, the weight of the image feature vector can be increased; if the user pays more attention to the content description of the advertisement, the weight of the text feature vector can be increased. In this way, the system can more flexibly process multi-modal data and generate advertisement feature sets that better meet user interests, thereby improving the quality of advertisement recommendation and user experience.

[0056] In some embodiments, the fusion process adopts an attention weighting mechanism, including: According to the user device type, dynamically adjusting the fusion weight coefficients of the text feature vector, the image feature vector, and the video feature vector; When the user device is a mobile terminal, the weight coefficient of the image feature vector is set to 1.2 times the baseline weight.

[0057] It should be noted that the present application adopts an attention weighting mechanism when generating the advertisement feature set by fusing the text feature vector, image feature vector and video feature vector. This mechanism can dynamically adjust the fusion weight coefficients of each feature vector according to the user device type. For example, when the user device is a mobile terminal, considering that mobile device users tend to quickly obtain information through images, the system will set the weight coefficient of the image feature vector to 1.2 times the baseline weight. This method can better adapt to the user needs of different devices and improve the accuracy of advertisement recommendation and user experience.

[0058] Specifically, the attention weighting mechanism is a method of dynamically adjusting the weight of feature vectors, which can change the importance of each feature vector in the fusion process according to different context information, such as user device type. The baseline weight here refers to the default weight of each feature vector when there is no specific context information. For example, the baseline weights of the text feature vector, image feature vector and video feature vector can be set to 1 respectively. When the user device is a mobile terminal, the system will adjust the weight of the image feature vector from 1 to 1.2, which means that the image feature vector will contribute more to the final advertisement feature set in the fusion process. This adjustment is based on the assumption that mobile device users tend to quickly obtain information through images. In this way, the system can more flexibly handle the user needs of different devices and improve the relevance of advertisement recommendation and user experience.

[0059] Preferably, when implementing the attention weighting mechanism, the system can further refine the weight adjustment strategy. For example, in addition to adjusting the weight according to the user device type, other factors such as the user's historical behavior data can also be considered. If the user has shown high attention to text content in the past behavior, the system can appropriately increase the weight of the text feature vector. In addition, when building the attention model, a deep learning framework such as TensorFlow or PyTorch can be used to train the model. The input parameters can include user device type, user's historical behavior data and each feature vector itself. The training goal of the model is to learn how to dynamically adjust the weight of each feature vector according to these inputs. When processing user feedback, the system can update the weight adjustment strategy in real time to adapt to changes in user interest. For example, if the user frequently clicks on image ads over a period of time, the system can dynamically increase the weight of the image feature vector. Through these optimization measures, the system can more accurately reflect the user's interest preferences and improve the quality of advertisement recommendation and user experience.

[0060] In some embodiments, after generating the advertisement material set, it further includes: monitoring the user's dwell time data and click behavior data on the advertisement material in real time; When the staying time of the continuous N advertisement materials is detected to be lower than the preset time threshold and there is no click behavior, a regeneration process of the candidate advertisement set is triggered; the regeneration process includes emptying the current candidate advertisement set and returning to performing the operation of determining the candidate advertisement set based on the user's advertisement preference screening information.

[0061] It should be noted that, after generating the advertisement material set, the present application will monitor the user's staying time data and click behavior data of the advertisement material in real time. When the staying time of the continuous N advertisement materials is detected to be lower than the preset time threshold and there is no click behavior, the system will trigger the regeneration process of the candidate advertisement set. This process includes emptying the current candidate advertisement set and returning to performing the operation of determining the candidate advertisement set based on the user's advertisement preference screening information. The staying time data here refers to the length of time that the user stays on each advertisement material, and the click behavior data refers to whether the user has performed a click operation on the advertisement material. By monitoring these data, the system can evaluate the user's interest degree in the current advertisement material, and when the user shows low interest in multiple advertisement materials, the candidate advertisement set is regenerated to provide advertisement recommendations that are more in line with the user's interest.

[0062] Specifically, the real-time monitoring of the user's staying time data and click behavior data of the advertisement material is realized by embedding data tracking code on the advertisement display interface. When the user browses the advertisement material, the system will record the display time of each advertisement material and whether the user has performed a click operation. The staying time data can be obtained by calculating the difference between the start time and the end time of the advertisement material display, and the click behavior data is recorded by monitoring the click event of the user on the advertisement material. The preset time threshold is a parameter determined by experience or experiment, which is used to judge whether the user is interested in the advertisement material. For example, if the preset time threshold is 5 seconds, when the user stays on a certain advertisement material for less than 5 seconds and has no click behavior, the system will consider that the user is not interested in the advertisement material. The parameter N of the continuous N advertisement materials can be set according to the frequency of advertisement display and the behavior pattern of the user, for example, N can be set to 3 or 5, indicating that when the user shows low interest in 3 or 5 continuous advertisement materials, the operation of regenerating the candidate advertisement set is triggered. This operation includes emptying the current candidate advertisement set and determining a new candidate advertisement set based on the user's advertisement preference screening information to update the advertisement recommendation content.

[0063] Preferably, when implementing the process of regenerating the candidate advertisement set, the system can further optimize the monitoring and processing logic. For example, in addition to monitoring the dwell time and click behavior, other user behavior data such as the number of swipes, zoom-in views, etc. can also be considered to more comprehensively evaluate user interest. In building the monitoring model, machine learning algorithms such as decision trees or random forests can be used to analyze user behavior data and predict user interest. Input parameters can include user dwell time, number of clicks, number of swipes, etc. The training goal of the model is to learn how to judge whether the user is interested in the advertisement material based on these behavior data. When processing user feedback, the system can update the user interest model in real time to adapt to changes in user interest. For example, if a user frequently clicks on a certain type of advertisement over a period of time, the system can dynamically adjust the advertisement preference filtering information to increase the proportion of that type of advertisement in the candidate advertisement set. Through these optimization measures, the system can more accurately capture changes in user interest and update the advertisement recommendation content in a timely manner, improving user experience and advertisement effectiveness.

[0064] The above-mentioned various embodiments of the present application have the following beneficial effects: 1. By combining user-initiated advertisement identification information and historical interaction records to generate a candidate advertisement set, the problem of incomplete interest coverage caused by traditional recommendation systems relying on single behavior data is solved, both explicit preference needs of users are met and potential interest is mined, and the relevance of candidate advertisements is improved. 2. Adopting a local and cloud collaborative flow segmentation processing mechanism, the system quickly matches advertisement materials based on multi-dimensional feature information in real-time advertisement streams, solves the problem of low advertisement retrieval efficiency in high concurrency scenarios, and through deduplication processing avoids repeated content display, optimizing the accuracy of advertisement resource allocation. 3. Introducing multi-modal feature fusion and dynamic weight adjustment technology, the system optimizes the fusion strategy of text, image and video features for different terminal devices, solves the problem of poor adaptability of cross-media advertisement materials, makes the recommendation results more consistent with user device characteristics and interaction habits, and improves the advertisement presentation effect.

[0065] Further, the storage medium of the embodiments of the present application stores program instructions capable of realizing all the methods described above, wherein the program instructions can be stored in the storage medium in the form of a software product, and include a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor execute all or part of the steps of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes, or a computer, a server, a mobile phone, a tablet, and other terminal devices.

[0066] The above description is merely some preferred embodiments of the present application and a description of the principles of the technology used. Those skilled in the art should understand that the scope of the application involved in the embodiments of the present application is not limited to the technical solutions formed by the specific combinations of the technical features described above, and should also cover other technical solutions formed by any combinations of the technical features described above or their equivalent features without departing from the inventive concept described above. For example, the technical solutions formed by mutually replacing the above-described features and the technical features disclosed in the embodiments of the present application (but not limited to) having similar functions.

Claims

1. An artificial intelligence-based advertising recommendation method, characterized in that: include: Determining a candidate advertisement set for the user based on the user's advertisement preference screening information, wherein the candidate advertisement set includes candidate advertisements that the user is interested in; The advertisement preference screening information is advertisement identification information input by the user in the advertisement selection interface, or is the user's historical advertisement interaction record; Obtaining feature information of each candidate advertisement in the candidate advertisement set; Based on the characteristic information of each candidate advertisement, a stream segmentation process is performed on the real-time advertisement stream of the advertisement delivery platform to obtain an advertisement material set of each candidate advertisement.

2. The method according to claim 1, characterized in that The step of filtering the user's advertisement preference information to determine the user's candidate advertisement set includes: When the advertisement preference screening information is advertisement identification information input by the user in an advertisement selection interface, generating a candidate advertisement set based on each advertisement identification in the advertisement identification information; When the advertisement preference screening information is the user's historical advertisement interaction record, obtaining the user's preferred advertisement sequence from a historical advertisement library based on the historical advertisement interaction record; Each advertisement element in the preferred advertisement sequence is identified, candidate advertisements are screened from each advertisement element, and a candidate advertisement set is generated based on the screened candidate advertisements.

3. The method according to claim 2, characterized in that The screening of candidate advertisements from each of the advertisement elements includes: For each of the advertisement elements, determining a weight value of the advertisement element based on the historical advertisement interaction record and the preferred advertisement sequence, and determining an estimated interest value of the user in the advertisement element based on the weight value of the advertisement element; A candidate advertisement is selected from each of the advertisement elements based on the estimated interest value of each of the advertisement elements.

4. The method according to claim 1, wherein The step of performing stream segmentation processing on the real-time advertisement stream of the advertisement delivery platform based on the characteristic information of each candidate advertisement to obtain a set of advertisement materials for each candidate advertisement includes: For each candidate advertisement, extracting first advertisement materials including the candidate advertisement from the real-time advertisement stream based on feature information of the candidate advertisement; Uploading the characteristic information of each candidate advertisement and the real-time advertisement stream of the advertisement delivery platform to the recommendation server to obtain cloud-based advertisement data of each candidate advertisement fed back by the recommendation server, wherein the cloud-based advertisement data of each candidate advertisement includes each second advertisement material corresponding to the characteristic information of the candidate advertisement extracted by the recommendation server from the real-time advertisement stream; For each candidate advertisement, deduplication processing is performed on each first advertisement material and each second advertisement material of the candidate advertisement to obtain each deduplicated advertisement material, and an advertisement material set of the candidate advertisement is generated based on each deduplicated advertisement material.

5. The method according to claim 4, characterized in that Generating the advertising material set of the candidate advertisement based on the deduplicated advertising materials includes: Determine each of the removed duplicate creatives as a target creative; Based on a preset advertisement grouping rule, the target advertisement materials are grouped to obtain a plurality of advertisement material groups, each of which contains at least one target advertisement material.

6. The method according to claim 5, characterized in that Also includes: Displaying the advertising material set of each candidate advertisement to the user; When receiving the trigger command fed back by the user, determining a target advertising material set from among the advertising material sets based on the trigger command; Determine a target advertising creative group in the target advertising creative set using the trigger command; Outputting each advertising creative in the target advertising creative group to the user.

7. The method according to claim 4, characterized in that The performing deduplication processing on each first advertising material and each second advertising material of the candidate advertisement includes: Extracting a first feature vector of each first advertising material and a second feature vector of each second advertising material; Calculating the similarity between each first eigenvector and all second eigenvectors; When there is a feature vector pair whose similarity value exceeds a preset threshold, the lower priority creative in the feature vector pair is deleted.

8. The method according to claim 1, characterized in that The step of obtaining characteristic information of candidate advertisements includes: Parsing multimodal data of the candidate advertisement, the multimodal data including text description data, image pixel data, and video frame sequence data; Generate text feature vectors, image feature vectors, and video feature vectors respectively through the pre-trained feature extraction model; The text feature vector, image feature vector and video feature vector are integrated to generate an advertisement feature set.

9. The method according to claim 8, characterized in that The fusion process adopts an attention weighting mechanism, including: Dynamically adjust the fusion weight coefficients of text feature vectors, image feature vectors, and video feature vectors according to the user device type; When the user equipment is a mobile terminal, the weight coefficient of the image feature vector is set to 1.2 times the reference weight.

10. The method according to claim 1, characterized in that When generating a creative bundle, also include: Real-time monitoring of user dwell time and click behavior data on advertising materials; When it is detected that the dwell time of N consecutive advertising materials is lower than the preset time threshold and there is no click behavior, the regeneration process of the candidate advertising set is triggered; the regeneration process includes clearing the current candidate advertising set and returning to perform the operation of determining the candidate advertising set based on the user's advertising preference screening information.

Images