Advertisement user analysis method and system

Abstract

The present invention discloses an advertising user analysis method and system, comprising: first, obtaining user behavior data in real time by responding to advertising promotion operations, and inputting the data into a trained advertising user behavior recognition model for classification. According to the classification results, the system determines a personalized push strategy from an advertising push strategy library, thereby achieving accurate analysis of user behavior and effective advertising push. Such a design improves the accuracy of advertising push and user satisfaction.

Description

Technical Field

[0001] The present invention relates to the field of advertisement push technology, and in particular to an advertisement user analysis method and system. Background Art

[0002] In existing advertising push systems, advertising user analysis is often based on simple user portraits and static behavioral data, which makes it difficult to accurately grasp users' dynamic interests and needs, resulting in inefficient advertising push. Summary of the invention

[0003] The object of the present invention is to provide an advertisement user analysis method and system.

[0004] In a first aspect, an embodiment of the present invention provides an advertisement user analysis method, the method comprising:

[0005] In response to an advertising promotion operation triggered by a current user, obtaining a current user behavior of the current user in a preset monitoring period;

[0006] Transmitting the current user behavior to the trained advertising user behavior recognition model, and obtaining a classification result of the current user behavior through the trained advertising user behavior recognition model;

[0007] According to the classification result, determining the advertisement push strategy corresponding to the current user from a pre-stored advertisement push strategy library;

[0008] The classification result and the advertisement push strategy are used as the advertisement user analysis result of the current user.

[0009] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is used to execute the method described in the first aspect.

[0010] Compared with the prior art, the beneficial effects provided by the present invention include: using an advertising user analysis method and system disclosed by the present invention, including: first, obtaining user behavior data in real time by responding to advertising promotion operations, and inputting it into the trained advertising user behavior recognition model for classification. According to the classification results, the system determines a personalized push strategy from the advertising push strategy library, thereby achieving accurate analysis of user behavior and effective push of advertisements. Such a design improves the accuracy of advertising push and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0012] Figure 1 A schematic diagram of the steps of the advertising user analysis method provided by an embodiment of the present invention;

[0013] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] 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. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0015] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0016] In order to solve the technical problems in the aforementioned background technology, Figure 1 This is a flow chart of the advertising user analysis method provided in an embodiment of the present disclosure. The advertising user analysis method is introduced in detail below.

[0017] Step S201, in response to an advertising promotion operation triggered by a current user, obtaining a current user behavior of the current user in a preset monitoring period;

[0018] Step S202, transmitting the current user behavior to the trained advertising user behavior recognition model, and obtaining a classification result of the current user behavior through the trained advertising user behavior recognition model;

[0019] Step S203, determining the advertisement push strategy corresponding to the current user from a pre-stored advertisement push strategy library according to the classification result;

[0020] Step S204: Using the classification result and the advertisement push strategy as the advertisement user analysis result of the current user.

[0021] In an embodiment of the present invention, illustratively, the server detects that user A clicks an advertising link on a social media application, and this operation is identified as a trigger for an advertising promotion operation. Then, the server begins to monitor the behavior of user A in the next preset monitoring period (e.g., the past week). These behaviors include, but are not limited to: user A's browsing history, search history, likes and comments on the social media application, etc. User A clicks on an advertising link about a "new smartphone", and the server begins to record user A's behavior in the next 7 days. During these 7 days, the server collects that user A browsed multiple articles related to mobile phones, entered keywords such as "mobile phone review" in the search bar, liked several comments about the camera function of mobile phones, etc. The server inputs the collected behavior data of user A into the trained advertising user behavior recognition model. This model is trained with a large amount of data and can recognize the user's behavior patterns and classify them into different behavior categories. For example, these categories can include "potential buyers", "interested but hesitant", "unintentional buyers", etc. The server inputs the browsing history, search history, likes and comments of user A into the advertising user behavior recognition model. After calculation, the model classifies user A's behavior as a "potential buyer" because user A's behavior shows continued attention and interest in mobile phone products. The server searches for the corresponding advertising push strategy in the pre-stored advertising push strategy library based on the user behavior classification results given by the advertising user behavior recognition model. These strategies may include the content, frequency, and time of the pushed advertisements. Since user A is classified as a "potential buyer", the server finds the corresponding strategy in the advertising push strategy library and decides to push an advertisement for the latest smartphone to user A, and push it once a day in the next week, with the push time selected at the beginning of user A's active time period. The server saves the user behavior classification results and the corresponding advertising push strategy as the results of advertising user analysis, which can be used for subsequent advertising push and effect evaluation. The server saves the result that user A is classified as a "potential buyer" and the corresponding advertising push strategy. This result will be used to push advertisements to user A in the future, and can be used to evaluate the effect of this advertising push, such as click-through rate, conversion rate, etc.

[0022] In an embodiment of the present invention, the advertisement user behavior recognition model is obtained in the following manner.

[0023] Acquire a target behavior instance and at least two benchmark behavior categories, wherein the target behavior instance corresponds to at least two sample behavior categories;

[0024] Obtain a baseline behavior category vector for each of the baseline behavior categories through a behavior feature extraction model, perform aggregation operations on at least two of the baseline behavior categories according to the baseline behavior category vectors to obtain at least two baseline behavior category centers, and obtain an aggregated behavior category for each of the baseline behavior category centers through the sample behavior categories;

[0025] The target behavior instance is transmitted to an advertising user behavior recognition model, and an inferred behavior category center is obtained from at least two of the benchmark behavior category centers through the advertising user behavior recognition model, and the benchmark behavior category corresponding to the target behavior instance is obtained from the inferred behavior category center;

[0026] Obtaining a first deviation based on a determination result of the inferred behavior category center and the aggregated behavior category, and obtaining a second deviation based on a determination result of the benchmark behavior category corresponding to the target behavior instance and the sample behavior category;

[0027] According to the first deviation and the second deviation, an integrated training process is performed on the behavior feature extraction model and the advertising user behavior recognition model to obtain a trained advertising user behavior recognition model.

[0028] In an embodiment of the present invention, illustratively, the server first collects a series of target behavior instances, which are the actual behavior data of users during the advertising promotion process. At the same time, the server also defines multiple benchmark behavior categories, such as "active buyers", "occasional followers" and "potential lost users". For each target behavior instance, the server marks at least two sample behavior categories, which are based on expert judgment or preliminary analysis. The server collects the behavior data of user B during the advertising promotion as the target behavior instance. For user B, the expert preliminarily judges that his behavior may belong to the two sample behavior categories of "active buyers" or "occasional followers". The server uses the behavior feature extraction model to process the behavior data of the benchmark behavior category, extract key features and generate a benchmark behavior category vector. Then, these vectors are aggregated (such as calculating the average or cluster center) to obtain the center of each benchmark behavior category. Finally, based on the labeled sample behavior categories, the server further aggregates the benchmark behavior category center into a higher-level aggregated behavior category. For the two benchmark behavior categories of "active buyers" and "occasional followers", the server uses the behavior feature extraction model to extract their feature vectors respectively and calculate their respective category centers. Then, based on the sample behavior category labeling of user B, the server aggregates the two category centers into a more specific aggregate behavior category, such as "high-intention purchaser". The server transmits the target behavior instance (such as user B's behavior data) to the advertising user behavior recognition model. The model infers the most likely behavior category center based on the similarity between the characteristics of the target behavior instance and the benchmark behavior category center, and determines the benchmark behavior category corresponding to the target behavior instance accordingly. The server inputs the behavior data of user B into the advertising user behavior recognition model. Based on the similarity between the behavioral characteristics of user B and the category centers of "active purchaser" and "occasional follower", the model infers that user B's behavior is closer to the category center of "active purchaser", so user B is classified as "active purchaser". The server compares the judgment result of the inferred behavior category center of the advertising user behavior recognition model with the actual value of the aggregate behavior category, and calculates the first deviation. At the same time, the server also compares the model's judgment result of the benchmark behavior category of the target behavior instance with the actual value of the sample behavior category, and calculates the second deviation. The server discovered that the advertising user behavior recognition model judged user B as an "active buyer", while the actual value in the aggregated behavior category was "high-intention buyer". There was a difference between the two, resulting in the first deviation. At the same time, the model judged user B as an "active buyer", while the label of the sample behavior category was "active buyer" or "occasional follower". If the label was "occasional follower", a second deviation would be generated. The server uses the first and second deviations as feedback to jointly adjust and optimize the behavior feature extraction model and the advertising user behavior recognition model. This process is an iterative process that improves the accuracy and generalization ability of the model by continuously reducing the deviation.The server adjusts and optimizes the parameters of the behavior feature extraction model and the advertising user behavior recognition model according to the first deviation and the second deviation. After multiple iterations of training, the deviation of the model gradually decreases, and finally a trained advertising user behavior recognition model is obtained, which can more accurately identify the user's behavior category.

[0029] In an embodiment of the present invention, the advertising user behavior recognition model includes a feature extraction component and a feature recognition component; the transmission of the target behavior instance to the advertising user behavior recognition model, and obtaining the inferred behavior category center from at least two of the benchmark behavior category centers through the advertising user behavior recognition model can be implemented through the following example.

[0030] Transmitting the target behavior instance to the feature extraction component to obtain an instance feature vector;

[0031] Transmitting the instance feature vector to the feature recognition component to obtain a first category reliability coefficient of each of the benchmark behavior category centers;

[0032] Performing a standardization operation on each of the first category reliability coefficients to obtain the first category belonging confidence of each of the reference behavior category centers;

[0033] When the first category attribution confidence is not lower than a preset first confidence threshold, the reference behavior category center corresponding to the first category attribution confidence is used as the inferred behavior category center.

[0034] In an embodiment of the present invention, illustratively, after the server collects the target behavior instance of the user, it will transmit these behavior data (such as the user's browsing history, search history, likes and comments, etc.) to the feature extraction component in the advertising user behavior recognition model. The feature extraction component will process these behavior data, extract key features, and generate an instance feature vector. The server collects the behavior data of user C during the advertising promotion as the target behavior instance. These data include user C browsing multiple tourism-related articles, searching for keywords such as "travel destination recommendation", and liking several comments about travel experience. The server transmits these data to the feature extraction component, and the component extracts key features (such as the number of tourism-themed articles browsed by the user, the frequency of tourism-related keywords searched, the number of tourism comments liked, etc.), and generates an instance feature vector containing these features. After obtaining the instance feature vector, the server will transmit it to the feature recognition component in the advertising user behavior recognition model. The feature recognition component will calculate the reliability coefficient (also known as matching degree or similarity) of each benchmark behavior category for the target behavior instance based on the similarity between the vector and the pre-defined benchmark behavior category center. The server transmits the instance feature vector of user C to the feature recognition component. The component compares the similarity between this vector and the centers of the three benchmark behavior categories of "travel enthusiasts", "leisure tourists" and "planned travelers", and calculates the reliability coefficient of each benchmark behavior category for the behavior instance of user C. Assume that the calculated reliability coefficients are 0.85, 0.60 and 0.75 respectively. Since the reliability coefficients of the centers of different benchmark behavior categories may be in different numerical ranges, in order to compare and judge, the server will perform a standardization operation on these reliability coefficients and convert them into confidences within the same numerical range. The server performs a standardization operation on the calculated reliability coefficients. Assume that the standardization operation converts all reliability coefficients into values ​​between 0 and 1, and the larger the value, the higher the confidence. After standardization, the confidence of user C for the three benchmark behavior categories of "travel enthusiasts", "leisure tourists" and "planned travelers" are 0.90 (corresponding to a reliability coefficient of 0.85), 0.65 (corresponding to a reliability coefficient of 0.60) and 0.80 (corresponding to a reliability coefficient of 0.75). In order to determine which benchmark behavior category the target behavior instance is most likely to belong to, the server will set a confidence threshold. Only when the confidence of a certain benchmark behavior category reaches or exceeds this threshold, the server will regard it as the inferred behavior category center. The server sets a first confidence threshold of 0.75. After comparing the confidence of user C for each benchmark behavior category, it is found that the confidence of "travel enthusiasts" and "planning travelers" (0.90 and 0.80 respectively) both meet the threshold requirements, while the confidence of "leisure tourists" (0.65) is below the threshold.Therefore, the server uses the benchmark behavior category center of "travel enthusiast" with the highest confidence as the inferred behavior category center, and believes that user C most likely belongs to the behavior category of "travel enthusiast".

[0035] In an embodiment of the present invention, the determination result of the inferred behavior category center includes the first category belonging confidence of each of the benchmark behavior category centers; the first deviation is obtained based on the determination result of the inferred behavior category center and the aggregated behavior category, which can be implemented through the following example.

[0036] According to the aggregated behavior category of each of the benchmark behavior category centers, obtaining the aggregated behavior confidence of each of the benchmark behavior category centers;

[0037] Calculating the mean square error deviation between each of the aggregated behavior confidences and the corresponding first category attribution confidence to obtain at least two confidence deviations;

[0038] The sum of all the confidence deviations is taken as the first deviation.

[0039] In an embodiment of the present invention, illustratively, in the previous steps, the aggregated behavior category of each benchmark behavior category center has been obtained. In order to calculate the first deviation, the server needs to assign an aggregated behavior confidence to each benchmark behavior category center. This confidence can be set based on expert knowledge, historical data or other standards. Assume that there are three benchmark behavior category centers, namely "travel enthusiasts", "leisure tourists" and "planned travelers". According to the judgment of experts, aggregated behavior confidences are assigned to these three category centers, which are 0.9 (corresponding to "travel enthusiasts"), 0.7 (corresponding to "leisure tourists") and 0.8 (corresponding to "planned travelers"). These confidences reflect the reliability or importance of each aggregated behavior category. In the above process, the first category attribution confidence of each benchmark behavior category center for the target behavior instance has been obtained. Now, the server needs to calculate the deviation between the aggregated behavior confidence of each benchmark behavior category center and the corresponding first category attribution confidence. This deviation can be calculated by the mean square error (or other statistical measures). For the behavior instance of user C, the first category attribution confidences for "travel enthusiasts", "leisure tourists" and "planned travelers" are 0.90, 0.65 and 0.80 respectively (as shown in the previous example). At the same time, it is known that the aggregated behavior confidences of the three benchmark behavior category centers are 0.9, 0.7 and 0.8 respectively. Now, the server calculates the mean square deviation between the aggregated behavior confidence of each benchmark behavior category center and the corresponding first category attribution confidence. For example, for the "travel enthusiasts" category, the deviation is ((0.9-0.90)^2=0); for the "leisure tourists" category, the deviation is ((0.7-0.65)^2=0.0025); for the "planned travelers" category, the deviation is ((0.8-0.80)^2=0). After calculating the confidence deviation of each benchmark behavior category center, the server adds these deviation values ​​to obtain the total value of the first deviation. This total value reflects the overall degree of difference between the inferred behavior category center and the aggregated behavior category of the advertising user behavior recognition model. Add the confidence deviations of the three benchmark behavior category centers calculated above to get the total value of the first deviation: (0+0.0025+0=0.0025). This value represents the slight difference between the inference result of the behavior instance of user C in the advertising user behavior recognition model and the aggregated behavior category defined by experts or historical data. If the value of the first deviation is large, it means that the model may need further adjustment or optimization.

[0040] In the embodiment of the present invention, the step of obtaining the benchmark behavior category corresponding to the target behavior instance from the inferred behavior category center may be implemented through the following example.

[0041] Traversing each of the benchmark behavior categories in the inferred behavior category center, and calculating a vector distance between the instance feature vector and a benchmark behavior category vector of the benchmark behavior category;

[0042] Performing a normalization operation on the vector distance to obtain a second category belonging confidence of the benchmark behavior category;

[0043] When the second category attribution confidence is not lower than a preset second confidence threshold, the benchmark behavior category corresponding to the second category attribution confidence is used as the benchmark behavior category corresponding to the target behavior instance.

[0044] In an embodiment of the present invention, illustratively, after determining the inferred behavior category center through the feature recognition component, the server will further traverse the benchmark behavior categories represented by these centers. For each benchmark behavior category, the server will calculate the vector distance between the instance feature vector of the target behavior instance and the benchmark behavior category vector of the benchmark behavior category. This distance usually reflects the similarity or difference between the instance feature vector and the benchmark behavior category vector. Assume that the inferred behavior category center includes two benchmark behavior categories of "travel enthusiasts" and "planned travelers". The server will calculate the vector distance between the instance feature vector of user C and the benchmark behavior category vector of "travel enthusiasts", as well as the vector distance between the instance feature vector of user C and the benchmark behavior category vector of "planned travelers". These vector distances can be calculated by Euclidean distance, cosine similarity or other vector similarity measurement methods. Since the vector distances between different benchmark behavior categories may be in different numerical ranges, in order to compare and judge, the server needs to perform standardization operations on these vector distances and convert them into confidences within the same numerical range. This confidence reflects the possibility that the target behavior instance belongs to a certain benchmark behavior category. The server uses a certain standardization method (such as Min-Max normalization) to normalize the calculated vector distance. Assume that after normalization, the second category attribution confidence of the benchmark behavior category of "travel enthusiasts" is 0.85, and the second category attribution confidence of the benchmark behavior category of "planned travelers" is 0.70. These values ​​represent the similarity between the behavior instance of user C and these two benchmark behavior categories. In order to ensure the accuracy and reliability of the results, the server sets a second confidence threshold. Only when the second category attribution confidence of a benchmark behavior category reaches or exceeds this threshold, the server will use the benchmark behavior category as the benchmark behavior category corresponding to the target behavior instance. The server sets the second confidence threshold to 0.75. After comparing the second category attribution confidences of the two benchmark behavior categories of "travel enthusiasts" and "planned travelers", it is found that the confidence of "travel enthusiasts" (0.85) meets the threshold requirement, while the confidence of "planned travelers" (0.70) is lower than the threshold. Therefore, the server uses "travel enthusiasts" as the benchmark behavior category corresponding to the target behavior instance of user C.

[0045] In an embodiment of the present invention, the determination result of the benchmark behavior category corresponding to the target behavior instance includes the second category attribution confidence of each benchmark behavior category in the inferred behavior category center; the second deviation is obtained based on the determination result of the benchmark behavior category corresponding to the target behavior instance and the sample behavior category, which can be implemented through the following example.

[0046] According to the sample behavior category, obtaining the target category confidence of each of the benchmark behavior categories in the inferred behavior category center;

[0047] Calculating the mean square error deviation between each of the target category confidences and the corresponding second category belonging confidences to obtain at least two category deviations;

[0048] The sum of all the category deviations is taken as the second deviation.

[0049] In an embodiment of the present invention, illustratively, the server has some sample behavior data, which have been marked as specific behavior categories (i.e., sample behavior categories). The server will use these sample data to determine the target category confidence of each benchmark behavior category in the inferred behavior category center. This usually involves statistics and classification of sample data to calculate the frequency or probability of occurrence of each benchmark behavior category in the sample data. Assume that there are a large number of sample behavior data, which have been marked as three benchmark behavior categories of "travel enthusiasts", "leisure tourists" and "planned travelers". The server will count the number of occurrences or frequencies of each benchmark behavior category in these sample data and calculate the corresponding target category confidence. For example, among 1000 sample data, 300 are marked as "travel enthusiasts", 200 are marked as "leisure tourists", and 500 are marked as "planned travelers". Therefore, the target category confidences are 0.3 (corresponding to "travel enthusiasts"), 0.2 (corresponding to "leisure tourists") and 0.5 (corresponding to "planned travelers"). After determining the target category confidence for each benchmark behavior category, the server calculates the deviation between these probabilities and the second category attribution confidence of the corresponding benchmark behavior category. This deviation is usually calculated by the mean square error (or other statistical measure) to quantify the degree of difference between the target category confidence and the second category attribution confidence. Assume that the previously calculated second category attribution confidence of "travel enthusiasts" is 0.85 (as described in the previous step), and the corresponding target category confidence is 0.3. The server calculates the mean square error deviation between these two values. Similarly, the mean square error deviation between the target category confidence and the second category attribution confidence of the benchmark behavior categories of "leisure tourists" and "planned travelers" is also calculated. The server sums the calculated category deviations for each benchmark behavior category to obtain a total deviation value, namely the second deviation. This second deviation reflects the overall degree of difference between the second category attribution confidence of the benchmark behavior category in the inferred behavior category center and the target category confidence of the sample behavior category. Assume that the category deviation of "travel enthusiasts" is 0.05, the category deviation of "leisure tourists" is 0.03, and the category deviation of "planned travelers" is 0.02 (these values ​​are only examples, and the actual calculated values ​​will depend on the specific algorithm and data). The server adds these category deviations and obtains a total value of the second deviation of 0.05+0.03+0.02=0.1. This value represents the overall difference between the second category attribution confidence of the benchmark behavior category in the inferred behavior category center and the target category confidence of the sample behavior category. If the value of the second deviation is large, it may mean that the advertising user behavior recognition model needs further adjustment or optimization.

[0050] In the embodiment of the present invention, the transmitting of the target behavior instance to the feature extraction component to obtain the instance feature vector may be implemented through the following example.

[0051] Performing time-series segmentation processing on the target behavior instance to obtain a user behavior set, wherein the user behavior set includes at least two user interaction operations;

[0052] Configuring an initial tag for an initial interaction operation of the user behavior set, and configuring a termination tag for a termination interaction operation of the user behavior set, to obtain a to-be-processed behavior set;

[0053] Performing feature extraction processing on the set of behaviors to be processed to obtain a set of behavior operation vectors;

[0054] By means of a recurrent neural network, the feature extraction component is used to perform a feature extraction operation on the behavior operation vector set to obtain an interaction feature vector set, wherein the interaction feature vector set includes a feature representation of each user interaction operation in the behavior set to be processed;

[0055] Through a recurrent neural network, feature merging is performed on each of the feature representations to obtain an instance feature vector.

[0056] In an embodiment of the present invention, illustratively, the server receives a target behavior instance, which can be a record of a user's behavior within a time period. The server performs time-series segmentation processing on the behavior instance, dividing it into multiple independent user interaction operations, which are combined into a user behavior set in chronological order. Assume that the server receives a target behavior instance of user C on an e-commerce platform, which records all operations of user C from entering the platform to leaving the platform. The server segments the behavior instance into multiple user interaction operations, such as "browse product A", "add to shopping cart", "browse product B", "submit order", etc., according to the chronological order of user C's operations, and these operations constitute a user behavior set. The server assigns a label to each interaction operation in the user behavior set to identify the start and end of the operation. Usually, the first interaction operation will be marked as the start operation, and the last operation will be marked as the end operation. These labels are helpful for subsequent feature extraction and model training. In the behavior set of user C, "browse product A" is the first operation, and the server will configure an initial label for it, indicating that this is the beginning of the user behavior. "Submit order" is the last operation of the user behavior, and the server will configure a termination label for it, indicating the end of the user behavior. After such processing, the server obtains a set of pending behaviors with start and end tags. The server extracts features from each user interaction operation in the pending behavior set and converts it into a vector form that can be understood by the computer. These vectors usually contain key information of user interaction operations, such as operation type, operation time, and operation result. The server extracts features from each interaction operation of user C, converts "browsing product A" into a vector containing information such as product ID, browsing time, and browsing duration; and converts "add to shopping cart" into a vector containing information such as product ID, adding time, and shopping cart quantity. After such processing, the server obtains a set of behavior operation vectors. The server uses a recurrent neural network (RNN) to process the set of behavior operation vectors. RNN can capture the temporal dependency in sequence data, thereby extracting features between user interaction operations. Through RNN processing, the server obtains a set of interaction feature vectors, which contains the feature representation of each interaction operation in the user behavior sequence. The server inputs the set of behavior operation vectors into the RNN model, and the model processes each vector and outputs the corresponding interaction feature vector. These feature vectors contain the temporal relationship and dependency information between each interaction operation in the user behavior sequence. After obtaining the set of interaction feature vectors, the server uses RNN or other sequence processing models to merge these feature vectors again. The purpose of this process is to merge the feature representations of multiple interaction operations in the user behavior sequence into a unified instance feature vector for subsequent classification or prediction tasks.The server inputs the interaction feature vector set into another RNN model, which merges these vectors and outputs a single instance feature vector. This vector integrates the feature information of all interaction operations in the user behavior sequence and can be used as a feature representation for subsequent model input.

[0057] In the embodiment of the present invention, the target behavior instance is obtained in the following manner.

[0058] Acquire a set of sample instances for training an advertisement user behavior recognition model, and determine a set of candidate instances to be screened based on the set of sample instances;

[0059] Acquire a multi-task integrated model for performing instance screening processing on the candidate instance set; the multi-task integrated model is a model composed of multiple sub-task networks matching multiple evaluation indicators, and one evaluation indicator corresponds to one sub-task network;

[0060] The candidate instance set is loaded into each subtask network of the multiple subtask networks, and each subtask network performs instance screening processing on the candidate instance set to obtain an instance screening result of the candidate instance set under each evaluation indicator;

[0061] Preprocessing the sample instance set based on the instance screening results of the candidate instance set under each evaluation indicator;

[0062] The target behavior instance is determined based on the preprocessed sample instance set.

[0063] In an embodiment of the present invention, illustratively, the server first retrieves a large amount of sample data related to advertising user behavior from a data warehouse, which contains various behavioral records of users on the advertising platform. Based on these sample data, the server applies some predefined rules or algorithms to filter out candidate instances with potential value as the basis for further screening and processing. Assume that the server obtains a set of sample instances of 100,000 user behaviors such as clicks, browsing, and purchases on the advertising platform. Based on these sample instances, the server uses data cleaning and screening algorithms to remove duplicate, abnormal, or insignificant instances, and finally determines a set of 50,000 candidate instances, which are more representative and relevant. The server then loads a pre-trained multi-task integration model. This model is composed of multiple sub-task networks, each of which corresponds to a specific evaluation indicator for evaluating the quality and value of candidate instances from different angles. The server loads a multi-task integration model containing three sub-task networks, corresponding to the three evaluation indicators of click-through rate (CTR), conversion rate (CVR), and user satisfaction (US). These three sub-task networks are trained to predict the performance of candidate instances on the corresponding evaluation indicators. The server inputs the candidate instance set into each subtask network in the multi-task ensemble model, and each subtask network will screen and score the candidate instances according to its own evaluation index. In this way, each candidate instance will get the screening results under multiple evaluation indicators. The server loads 50,000 candidate instances into three subtask networks. The first subtask network scores and ranks the candidate instances according to the click-through rate (CTR); the second subtask network scores and ranks the candidate instances according to the conversion rate (CVR); and the third subtask network scores and ranks according to user satisfaction (US). Finally, each candidate instance will get a screening result containing the scores of the three evaluation indicators. The server preprocesses the original sample instance set based on the screening results of the candidate instance set under multiple evaluation indicators. This can include deleting instances that do not meet the requirements, adjusting the weights of instances, etc., to ensure that the final target behavior instance set can better reflect the true situation of advertising user behavior. The server preprocesses the original 100,000 sample instance set based on the screening results of the three subtask networks. First, the server deletes instances that score too low under any evaluation indicator; then, different weights are assigned to each instance based on its score under different evaluation indicators. After such preprocessing, the quality of the sample instance set is improved. Finally, the server will determine the final target behavior instance based on the preprocessed sample instance set. These target behavior instances will be used as training data for the advertising user behavior recognition model to improve the accuracy and performance of the model. After preprocessing, the server screened out 20,000 high-quality target behavior instances from 100,000 original sample instances.These instances performed well in multiple evaluation indicators such as click-through rate, conversion rate, and user satisfaction, and were therefore selected as training data for the advertising user behavior recognition model.

[0064] In the embodiment of the present invention, the sample instance set includes a main sample instance set; and the determination of the candidate instance set to be screened based on the sample instance set can be implemented through the following examples.

[0065] Performing instance sampling on the main sample instance set to obtain sample instances matching the main sample instance set;

[0066] The sampled sample instance is used as a sample instance in a candidate instance set, and the candidate instance set to be screened is determined based on the sample instance in the candidate instance set.

[0067] In an embodiment of the present invention, the server stores a large amount of user behavior data, which are organized into multiple sample instance sets for different model training and data analysis tasks. Among them, the main sample instance set is a core data set containing a large amount of user behavior data, which is used for subsequent data screening and model training. The main sample instance set on the server contains all user behavior data that occurred on the advertising platform in the past year, totaling about 5 million. These data include user clicks, browsing, purchases, comments and other behaviors, as well as corresponding timestamps, advertising IDs, user IDs and other information. Since the amount of data in the main sample instance set may be very large, directly processing the entire set will consume a lot of computing resources and time. Therefore, the server will use sampling technology to extract a part of representative data from the main sample instance set as sampled sample instances. The server randomly samples the main sample instance set and extracts 100,000 pieces of data from it as sampled sample instances. These 100,000 pieces of data cover data from different time periods, different user groups and different advertising types, and can better represent the characteristics of the entire main sample instance set. The server uses the sampled sample instances as the initial data of the candidate instance set. Since there may be errors or deviations in the sampling process, the candidate instance set needs to be further screened and processed to ensure that the final data set has higher quality and representativeness. The server uses 100,000 sampled sample instances as the initial data for the candidate instance set. Then, the server removes duplicate data, abnormal data, and irrelevant data through a series of data cleaning and preprocessing operations to ensure the accuracy and consistency of the data. Then, the server further screens and sorts the candidate instance set according to business needs and data characteristics, and finally determines a set of 50,000 candidate instances to be screened. These instance sets have been reduced in number, but have been significantly improved in quality and representativeness, providing more reliable data support for subsequent model training and analysis.

[0068] In an embodiment of the present invention, the main sample instance set includes multiple sample instance data groups, and each sample instance data group in the multiple sample instance data groups includes at least one sample instance; the instance sampling of the main sample instance set to obtain the sampled sample instances matching the main sample instance set can be implemented through the following example.

[0069] Get the number of samples used for instance sampling;

[0070] Extracting sample instances corresponding to the sampling number from each of the sample instance data groups in sequence to obtain sample instance data groups corresponding to each of the sample instance data groups;

[0071] Based on the sampled sample instance data groups corresponding to the sample instance data groups, the sampled sample instances matching the main sample instance set are determined.

[0072] In an embodiment of the present invention, illustratively, in an advertising system, a server stores a large amount of user behavior data, which are organized into multiple sample instance data groups. Each sample instance data group represents a specific type of user behavior or advertising activity, such as click data of a certain advertising activity, browsing data of a certain time period, etc. The main sample instance set includes multiple sample instance data groups, such as a click data group of advertisement A, a purchase data group of advertisement B, a browsing data group of the first quarter, etc. Each data group contains multiple specific sample instances, such as a user click record of advertisement A, a user purchase record of advertisement B, etc. When deciding to perform sampling, the server needs to first determine a sampling number, which is usually determined based on a combination of factors such as business requirements, computing resource limitations, and data characteristics. Based on business requirements, the server decides to extract 1,000 sample instances from each sample instance data group as sampling samples. This number can ensure the representativeness of the sample without causing excessive computing pressure on the server. After determining the sampling number, the server begins to extract the corresponding number of sample instances from each sample instance data group in turn. The extraction process can be random or according to certain rules or strategies. The server randomly extracts 1,000 click records from the click data group of advertisement A, 1,000 purchase records from the purchase data group of advertisement B, and 1,000 browsing records from the browsing data group of the first quarter. These extracted sample instances constitute the sampling sample instance data group corresponding to each sample instance data group. After obtaining the sampling sample instance data group corresponding to each sample instance data group, the server needs to integrate them to form a sampling sample instance set that matches the main sample instance set. This set will serve as input data for subsequent data processing and model training. The server integrates 1,000 click records of advertisement A, 1,000 purchase records of advertisement B, and 1,000 browsing records of the first quarter into a sampling sample instance set containing 3,000 sample instances. This set is consistent with the main sample instance set in terms of data distribution, feature distribution, etc., and can therefore serve as a reliable input for subsequent data processing and model training.

[0073] In the embodiment of the present invention, the sample instance set includes a main sample instance set; the preprocessing of the sample instance set based on the instance screening result of the candidate instance set under each evaluation indicator can be implemented through the following examples.

[0074] Determining the evaluation indicators to be processed that need to be preprocessed based on the instance screening results of the candidate instance set under each evaluation indicator;

[0075] The sample instances in the main sample instance set are preprocessed based on the evaluation index to be processed.

[0076] In an embodiment of the present invention, illustratively, in an advertising data analysis system, a server stores a large amount of user behavior data, which are organized into a main sample instance set. This set contains detailed data on various aspects such as advertising activities and user behaviors, and is the basis for subsequent data analysis and model training. The main sample instance set contains user behavior data of all advertising activities in the past year, such as clicks, browsing, purchases, etc. Each piece of data records in detail information such as user ID, advertising ID, behavior type, timestamp, etc. The server analyzes which evaluation indicators have poor performance or abnormal data based on the screening results of the candidate instance set under multiple evaluation indicators, thereby determining the evaluation indicators that need to be preprocessed. The server inputs the candidate instance set into the multi-task integrated model and obtains the screening results under the three evaluation indicators of click-through rate (CTR), conversion rate (CVR) and user satisfaction (US). The analysis found that under the CVR indicator, the performance of the candidate instance set is generally poor, and there are a large number of outliers. Therefore, the server determines CVR as a pending evaluation indicator that needs to be preprocessed. After determining the pending evaluation indicator, the server will preprocess the sample instances in the main sample instance set for the indicator. The purpose of preprocessing is to improve data quality and reduce the impact of outliers on subsequent model training. Since CVR is the evaluation indicator to be processed, the server begins to perform CVR-related preprocessing on the sample instances in the main sample instance set. First, the server performs outlier detection on the CVR value and removes outliers that exceed the set threshold. Next, the server uses statistical methods (such as median filling, mean filling, etc.) to fill in the missing CVR values. Finally, the server also adjusts the weights of the sample instances according to the distribution of CVR values ​​to ensure the balance of data distribution. After these preprocessing steps, the data quality of the sample instances in the main sample instance set under the CVR indicator has been significantly improved.

[0077] In an embodiment of the present invention, the multiple evaluation indicators include a target evaluation indicator; the instance screening result under the target evaluation indicator includes the instance screening sub-result of each sample instance in the candidate instance set under the target evaluation indicator; the instance screening result based on the candidate instance set under each evaluation indicator determines the evaluation indicator to be processed that needs to be preprocessed, which can be implemented through the following examples.

[0078] In the candidate instance set, the sample instances whose instance screening sub-results under the target evaluation index are abnormal are taken as abnormal sample instances;

[0079] When the number of the abnormal sample instances reaches the preprocessing condition, the target evaluation index is used as a pending evaluation index that needs to be preprocessed.

[0080] In an embodiment of the present invention, exemplarily, in an advertising data analysis system, a server uses multiple evaluation indicators to measure advertising effects and user behaviors. Among them, a specific evaluation indicator is set as a "target evaluation indicator" because the indicator is crucial to business decisions or model performance. In advertising effect analysis, there are multiple evaluation indicators such as click-through rate (CTR), conversion rate (CVR), user satisfaction (US), etc. Among them, CVR (conversion rate) is set as a target evaluation indicator because the conversion rate is directly related to the economic benefits of advertising activities. The server screens each sample instance according to the performance of the candidate instance set under the target evaluation indicator and generates corresponding screening sub-results. These sub-results will be used for subsequent data analysis and preprocessing decisions. The server calculates the CVR value of each sample instance in the candidate instance set and determines whether the value is abnormal based on business rules or statistical methods. For example, if the CVR value of a sample instance is much lower than the historical average or industry average, it will be regarded as an abnormal state and a corresponding screening sub-result will be generated. The server identifies those sample instances with abnormal performance based on the screening sub-results under the target evaluation indicator and marks them as abnormal sample instances. These abnormal sample instances may be caused by data errors, abnormal behavior, or other reasons. In the advertising effectiveness analysis, the server found that the CVR values ​​of multiple sample instances in the candidate instance set were extremely low, far below the normal level. These sample instances were identified as abnormal sample instances by the server and marked. When the server detects that the number of abnormal sample instances reaches a preset threshold or meets other preprocessing conditions, it will consider that there is a problem with the data quality of the target evaluation indicator and the indicator needs to be preprocessed. In the advertising effectiveness analysis, the server sets a threshold. When the number of abnormal sample instances exceeds this threshold, the server will consider that there is a problem with the CVR data. Once this condition is met, the server will use CVR as a pending evaluation indicator that needs to be preprocessed, and start the corresponding preprocessing process, such as data cleaning, outlier processing, data reconstruction, etc. In this way, the server can ensure that high-quality data is used for subsequent data analysis and model training.

[0081] In the embodiment of the present invention, the sample instance set includes an auxiliary sample instance set; and the determination of the candidate instance set to be screened based on the sample instance set may be implemented through the following examples.

[0082] The sample instances in the auxiliary sample instance set are used as sample instances in the candidate instance set, and the candidate instance set to be screened is determined based on the sample instances in the candidate instance set.

[0083] In the embodiment of the present invention, exemplarily, in the data analysis system, the server usually has a large number of data sets, including a main sample instance set and an auxiliary sample instance set. The auxiliary sample instance set usually contains some additional, non-core data, which can be used to supplement the information of the main sample instance set and improve the accuracy and comprehensiveness of data analysis. In advertising data analysis, the main sample instance set contains core advertising behavior data such as user clicks, browsing and purchases. The auxiliary sample instance set can contain other relevant information of the user, such as the user's search history, browsing preferences, geographic location, etc. Although this information is not a core indicator for directly measuring the advertising effect, it can provide more contextual information to help the server understand user behavior and advertising effect more accurately. When determining the candidate instance set to be screened, the server will take the sample instances in the auxiliary sample instance set into consideration. This is because although the data in the auxiliary sample instance set may not be a core indicator, it can also contain some valuable information for data analysis. In advertising data analysis, the server decides to screen the candidate instance set before model training. It first uses all sample instances in the auxiliary sample instance set (such as the user's search history, browsing preferences, etc.) as the initial sample instances of the candidate instance set. These sample instances will be subsequently screened and processed together with other sample instances from the primary sample instance set. After taking the sample instances in the auxiliary sample instance set as the initial sample instances of the candidate instance set, the server will screen and process the candidate instance set based on certain rules or algorithms. This screening process may involve steps such as data cleaning, outlier detection, and data conversion, aiming to remove invalid or low-quality sample instances and retain high-quality and valuable sample instances. In the advertising data analysis, the server performs data cleaning and outlier detection on the candidate instance set. It first removes those sample instances that contain invalid data or lack key information (such as empty user ID, non-existent advertising ID, etc.). Then, it uses statistical methods to detect and remove outliers (such as CVR values ​​that are much higher or much lower than normal levels). After these screening steps, the server obtains a cleaned and optimized candidate instance set, which will be used as input data for subsequent model training and data analysis.

[0084] In an embodiment of the present invention, the sample instance set includes an auxiliary sample instance set; the multiple evaluation indicators include a target evaluation indicator; the instance screening result under the target evaluation indicator includes the instance screening sub-result of each sample instance in the candidate instance set under the target evaluation indicator; the preprocessing of the sample instance set based on the instance screening result of the candidate instance set under each evaluation indicator can be implemented through the following examples.

[0085] In the candidate instance set, the sample instances whose instance screening sub-results under the target evaluation index are abnormal are taken as abnormal sample instances;

[0086] The abnormal sample instance is preprocessed based on the target evaluation index.

[0087] In an embodiment of the present invention, exemplarily, in an advertising data analysis system, a server maintains a sample instance set including a main sample instance set and an auxiliary sample instance set. The auxiliary sample instance set contains additional information related to the advertising campaign, such as the user's browsing history, search history, etc., which may help to more accurately analyze the advertising effect. In the advertising data analysis system, the main sample instance set contains core data such as advertising clicks, displays, and conversions. The auxiliary sample instance set contains information such as the user's search keywords, browsed pages, and geographic locations. Although this information does not directly participate in the evaluation of advertising effects, it can be used as supplementary information to help the server understand user behavior more comprehensively. When analyzing advertising effects, the server uses multiple evaluation indicators to measure the performance of the advertisement. Among them, a specific evaluation indicator is set as a "target evaluation indicator" because the indicator is crucial to business decisions or model performance. In the advertising effect analysis, the server uses multiple evaluation indicators such as click-through rate (CTR), conversion rate (CVR), and user satisfaction (US). Among them, CVR (conversion rate) is set as the target evaluation indicator because the conversion rate is directly related to the economic benefits of the advertising campaign. The server screens each sample instance in the candidate instance set under the target evaluation index and generates corresponding screening sub-results. These sub-results will be used for subsequent data analysis and preprocessing decisions. The server calculates the CVR value of each sample instance in the candidate instance set and determines whether the value is abnormal based on the set threshold or business rules. If the CVR value of a sample instance is far below the normal level, it will be regarded as an abnormal state and a corresponding screening sub-result will be generated. The server identifies those sample instances that behave abnormally based on the screening sub-results under the target evaluation index and marks them as abnormal sample instances. These abnormal sample instances can be caused by data errors, abnormal behavior or other reasons. In the advertising effect analysis, the server found that the CVR values ​​of multiple sample instances in the candidate instance set were extremely low, far below the normal level. These sample instances were identified as abnormal sample instances by the server and marked. After the server identifies the abnormal sample instances, it will preprocess these abnormal sample instances based on the target evaluation index. The purpose of preprocessing is to improve data quality and reduce the impact of abnormal values ​​on subsequent model training or data analysis. For the abnormal sample instances with extremely low CVR values ​​identified above, the server can adopt a variety of preprocessing strategies, such as removing these sample instances (if they are considered invalid or misleading), or using more robust statistical methods (such as median filling) to replace these outliers. In this way, the server can ensure that subsequent data analysis and model training use more accurate and reliable data.

[0088] In an embodiment of the present invention, a target evaluation indicator among the multiple evaluation indicators includes multiple evaluation sub-indicators; a subtask network matching the target evaluation indicator is composed of multiple subtask units matching the multiple evaluation sub-indicators, and one evaluation sub-indicator corresponds to one subtask unit; the candidate instance set is loaded into each subtask network among the multiple subtask networks, and each subtask network performs instance screening processing on the candidate instance set to obtain the instance screening result of the candidate instance set under each evaluation indicator, which can be implemented through the following examples.

[0089] The candidate instance set is loaded into each subtask unit of the plurality of subtask units, and each subtask unit performs instance screening processing on the candidate instance set to obtain an instance screening result of the candidate instance set under each evaluation sub-indicator;

[0090] Based on the instance screening results under each of the evaluation sub-indicators, determining the instance screening results of the candidate instance set under the target evaluation indicator;

[0091] Based on the instance screening result of the candidate instance set under the target evaluation indicator, the instance screening result of the candidate instance set under each evaluation indicator is determined.

[0092] In an embodiment of the present invention, exemplarily, in an advertising data analysis system, the server not only uses multiple evaluation indicators to measure the advertising effect, but also some key evaluation indicators (such as target evaluation indicators) are further subdivided into multiple evaluation sub-indicators. This can more comprehensively evaluate the performance of advertisements in various aspects. In the advertising effect analysis, the target evaluation indicator "advertising quality" is subdivided into multiple evaluation sub-indicators, such as "advertising relevance" (measuring the degree of matching between advertisements and user search intentions), "advertising creativity" (measuring the attractiveness of advertising content) and "advertising loading speed" (measuring the speed of advertisement loading). The server constructs a subtask network that matches the target evaluation indicator. This network consists of multiple subtask units, each of which is responsible for processing the data of an evaluation sub-indicator. This division of labor allows each subtask unit to focus on processing its specific evaluation sub-indicator, thereby improving processing efficiency and accuracy. In the advertising data analysis system, the server constructs a subtask network that matches the target evaluation indicator of "advertising quality". This network includes three subtask units: one is responsible for processing "advertising relevance" data, one is responsible for processing "advertising creativity" data, and another is responsible for processing "advertising loading speed" data. The server loads the candidate instance set into each subtask unit in the subtask network. Each subtask unit receives a portion of the candidate instance data and performs instance screening processing according to its corresponding evaluation sub-indicator. The processing result will be used as the instance screening result under the evaluation sub-indicator. The server loads the candidate ad instance set into three subtask units. The first subtask unit receives the candidate ad instance data and screens it according to the "ad relevance" evaluation sub-indicator to find out the ads that do not match the user's search intent. The second subtask unit screens out ads with poor creativity according to the "ad creativity" evaluation sub-indicator. The third subtask unit screens out ads with too slow loading speed according to the "ad loading speed" evaluation sub-indicator. The server collects the instance screening results of each subtask unit under its own evaluation sub-indicator, and comprehensively determines the instance screening results of the candidate instance set under the target evaluation indicator based on these results. This process may involve weighted, aggregated or comprehensive analysis of the results of multiple evaluation sub-indicators. The server collects the screening results of the three subtask units under the evaluation sub-indicators of "ad relevance", "ad creativity" and "ad loading speed". It then combines these results according to certain weights or rules to derive an overall score or status for each candidate ad under the "ad quality" target evaluation metric. Based on this score or status, the server can determine which ads are high-quality and which ads need further optimization or adjustment. Although the target evaluation metric is one of multiple evaluation metrics, the server usually also needs to determine the screening results of the candidate instance set under other evaluation metrics. This process may involve combining or correlating the screening results under the target evaluation metric with the results under other evaluation metrics.After determining the screening results of the candidate ad instance set under the "ad quality" target evaluation indicator, the server also needs to consider other evaluation indicators such as "click-through rate" (CTR), "conversion rate" (CVR), etc. It can conduct a comprehensive analysis of the screening results of "ad quality" and the results of these evaluation indicators to obtain a more comprehensive screening result of the candidate ad instance set. In this way, the server can optimize the advertising strategy and improve the advertising effect based on the evaluation results of multiple dimensions.

[0093] In an embodiment of the present invention, the sample instance set includes sample instance data groups classified by multiple user groups; the sample instance data groups classified by multiple user groups include a first user group classification sample instance data group and a second user group classification sample instance data group under the first user group classification sample instance data group; the embodiment of the present invention also provides the following implementation manner.

[0094] Obtaining a user group association rule for associating the first user group classification sample instance data group with the second user group classification sample instance data group; the user group association rule is used to characterize the corresponding relationship between the first user group classification sample instance data group and the second user group classification sample instance data group;

[0095] Based on the user group association rule, summarizing the user group analysis results of the instance screening results of the candidate instance set under each evaluation indicator to obtain user group analysis data matching the second user group classification sample instance data group;

[0096] Based on the user group association rule, user group analysis results are summarized for the instance screening results of the candidate instance set under each evaluation indicator to obtain user group analysis data that matches the first user group classification sample instance data group.

[0097] In an embodiment of the present invention, exemplarily, in an advertising data analysis system, a sample instance set maintained by a server includes sample instance data groups of multiple user group classifications. These user group classifications are divided based on user characteristics, behaviors or interests, so as to more accurately place advertisements and analyze advertising effects. In an e-commerce advertising system, a sample instance set includes sample instance data groups of multiple user group classifications, such as "young female user group", "middle-aged male user group", "student user group", etc. Each user group classification sample instance data group includes historical data such as advertisement clicks, browsing, and purchases of the corresponding user group. In some cases, user group classifications may have a hierarchical structure, that is, a large user group classification can be further subdivided into smaller sub-user group classifications. This helps to analyze the behaviors and needs of different user groups more finely. In an e-commerce advertising system, "young female user group" is a large user group classification, which can be further subdivided into two sub-user group classifications of "young female fashion seekers" and "young female housewives". These two sub-user group classifications correspond to different shopping preferences and advertising needs, respectively. In order to associate the screening results of the candidate instance set with the user group classification, the server needs to obtain user group association rules. These rules define the correspondence between different user group classifications for user group analysis. In the e-commerce advertising system, the server obtains the association rules between the "young female user group" and the two sub-user groups of "young female fashion seekers" and "young female housewives". These rules can include some conditions or logic, such as age range, shopping preferences, browsing behavior, etc., to determine whether a user belongs to a sub-user group. After the candidate instance set is screened by each evaluation indicator, the server will associate the screening results with the user group classification according to the user group association rules, and summarize the user group analysis data matching the specific user group classification (such as the second user group classification). The server screens the candidate advertising instance set according to the evaluation indicators and obtains the performance of each advertisement under the evaluation indicators such as click-through rate and conversion rate. Then, it associates the screening results with the user group classification according to the user group association rules. For example, for the sub-user group of "young female fashion seekers", the server will summarize the performance data of all advertisements matching the sub-user group under the evaluation indicators to form a user group analysis data report for the sub-user group. The process is similar to the previous one, but this time the user group analysis results are summarized for the first user group classification (i.e., the larger user group classification). After completing the user group analysis for the two sub-user groups of "young female fashion seekers" and "young female housewives", the server also needs to classify and summarize the larger user group of "young female user group".It integrates the analysis data of the two sub-user groups according to the user group association rules to form a comprehensive user group analysis data report for the "young female user group". This report can help advertisers understand the advertising needs and performance of this user group more comprehensively.

[0098] In the embodiments of the present invention, the following implementation modes are also provided.

[0099] Based on the instance screening results of the candidate instance set under each evaluation indicator, determining quantitative indicator information at multiple multidimensional data analysis levels;

[0100] Displaying a visualization interaction interface; the visualization interaction interface includes visualization components corresponding to the quantitative indicator information at the multiple multi-dimensional data analysis levels.

[0101] In an embodiment of the present invention, illustratively, after completing the screening process of the candidate instance set (such as advertising instance) under various evaluation indicators (such as click-through rate, conversion rate, etc.), the server will further conduct an in-depth analysis of the screening results to determine the quantitative indicator information at multiple different dimensions or levels. These quantitative indicator information can help users understand the performance and characteristics of the candidate instances more comprehensively. Assume that the server screens a set of candidate advertising instances in an e-commerce advertising system and obtains the performance of these advertisements under evaluation indicators such as click-through rate and conversion rate. Then, the server will perform a multi-dimensional analysis of these screening results, such as analyzing the effect of the advertisement from different dimensions such as region, time, and user portrait. For example, the server can determine the click-through rate distribution of an advertisement in different provinces, or the fluctuation of the conversion rate in different time periods of a day. These multi-dimensional quantitative indicator information will be extracted for subsequent visual display. In order to enable users to intuitively understand the performance of candidate instances under multiple evaluation indicators and multi-dimensional quantitative indicator information, the server will generate a visual interactive interface and display this information in the interface. This interface usually contains various visualization components such as charts and images, and users can interact with the data through these components to further explore and analyze the data. In the e-commerce advertising system, the server generates a visual interactive interface, and displays the screening results of candidate ads under multiple evaluation indicators and multi-dimensional quantitative indicator information in the interface. The interface can contain multiple different visualization components, such as a bar chart to show the distribution of click-through rates of ads in different provinces, and a line chart to show the fluctuation of conversion rates of ads in different time periods of a day. Users can interact with these components by clicking, dragging, and other operations to view data details under different dimensions, so as to have a deeper understanding of the effects and characteristics of the ads. In this scenario, the server first determines the quantitative indicator information at multiple multi-dimensional data analysis levels based on the screening results of the candidate instance set under multiple evaluation indicators. Then, it generates a visual interactive interface and displays the visualization components corresponding to these quantitative indicator information in the interface. Users can intuitively understand the performance and characteristics of the candidate instances through this interface, and conduct further data analysis and exploration.

[0102] In an embodiment of the present invention, the multiple evaluation indicators include a target evaluation indicator, and the multiple subtask networks include a target subtask network that matches the target evaluation indicator; the embodiment of the present invention also provides the following implementation manner.

[0103] Acquire a sample screening instance that matches the target evaluation indicator; the sample screening instance is associated with a screening target value, and the screening target value is used to characterize the actual screening result of the sample screening instance under the target evaluation indicator;

[0104] Acquire an original subtask network that matches the target subtask network, load the sample screening instance into the original subtask network, perform instance screening processing on the sample screening instance through the original subtask network, and obtain an inferred screening result of the sample screening instance under the target evaluation index;

[0105] The original subtask network is tuned based on the real screening result and the inferred screening result, and the original subtask network after the tuning operation is used as the target subtask network.

[0106] In an embodiment of the present invention, exemplarily, in a recommendation system, the server maintains multiple evaluation indicators, such as click-through rate, conversion rate, user satisfaction, etc. Among them, the conversion rate is determined as the target evaluation indicator because it is directly related to the core goal of the business - increasing sales. Multiple subtask networks are deployed in the server, and each subtask network is optimized for a specific evaluation indicator. The target subtask network that matches the conversion rate is selected for subsequent training and tuning. In the recommendation system of the e-commerce platform, the server has a subtask network called "conversion rate prediction", which is specifically used to predict the conversion rate of recommended products purchased by users. This subtask network is selected as the target subtask network that matches the target evaluation indicator of conversion rate. The server retrieves historical data related to the target evaluation indicator (conversion rate) from the data warehouse, and these data are called sample screening instances. Each sample screening instance is associated with a screening target value, which represents the actual screening result of the instance under the target evaluation indicator (conversion rate), that is, whether the user actually purchased the recommended product. The server retrieves the interaction data between the user and the recommended product in the past period of time from the database, and these data include user ID, recommended product ID, whether to buy (i.e., screening target value) and other information. Among them, the "whether to buy" field is the screening target value, which indicates whether the user actually bought the recommended product. The server loads the obtained sample screening instance into the original subtask network (i.e., the conversion rate prediction subtask network) that matches the target evaluation index. The original subtask network will perform instance screening on these sample screening instances, that is, predict the possible results of each instance under the target evaluation index (conversion rate). This prediction result is called the inferred screening result. The server inputs the retrieved interaction data between the user and the recommended product into the "conversion rate prediction" subtask network. The network processes and analyzes this data based on its learned patterns and rules, and then outputs the probability of each user purchasing the recommended product (i.e., the inferred screening result). The server compares the inferred screening result (predicted conversion rate) output by the original subtask network with the actual screening result (actual conversion rate) in the sample screening instance. Based on the comparison results, the server will adjust and optimize the parameters of the original subtask network to improve its prediction accuracy. This tuned subtask network will be used as the new target subtask network. The server compares the predicted conversion rate output by the "conversion rate prediction" subtask network with the actual conversion rate recorded in the database. If a large deviation is found between the predicted result and the actual result, the server will adjust the parameters of the network (such as weights, biases, etc.) to improve its prediction accuracy. After multiple iterations and optimizations, a "conversion rate prediction" subtask network with better performance is obtained, which will be used as the new target subtask network in subsequent recommendation tasks.

[0107] In the embodiment of the present invention, the sample instance set includes a main sample instance set, and the main sample instance set is stored in a sample instance data pool; the embodiment of the present invention also provides the following implementation manner.

[0108] When the preprocessed main sample instance set is determined, the main sample instance set stored in the sample instance data pool is optimized by using the preprocessed main sample instance set.

[0109] In the embodiment of the present invention, the server first collects a large amount of advertising user behavior data, which includes the user's click, browse, purchase and other behaviors, as well as various attribute information related to these behaviors, such as advertising content, time, geographic location, etc. Among these data, the server identifies some particularly important or representative sample instances, which are classified as "main sample instance sets". In order to facilitate subsequent use and management, the server saves these main sample instance sets in a specific data structure, called "sample instance data pool". For example, in an e-commerce advertising scenario, the server collects a large amount of data on users browsing and purchasing goods. Among them, some users are considered important users because of their high purchase frequency, large purchase amount or rich purchase variety. The browsing and purchase records of these important users are regarded as main sample instances and saved in the sample instance data pool. Before using the main sample instance set for advertising user analysis, the server will first perform preprocessing operations on these sample instances. The preprocessing operation may include steps such as data cleaning (such as removing duplicate data, filling missing values, etc.), feature extraction (such as extracting user interests, purchasing power and other features from user behavior data) and feature encoding (such as converting text data into numerical data). Taking the e-commerce advertising scenario as an example, the server will preprocess the user data in the main sample instance set. First, the server will remove duplicate user purchase records to ensure that each user's data is calculated only once. Then, the server will extract the user's purchase frequency, purchase amount, purchase type and other features for subsequent user analysis. Finally, the server will encode these features and convert them into numerical data suitable for machine learning model processing. After determining the preprocessed main sample instance set, the server will use these preprocessed sample instances to optimize the original main sample instance set in the sample instance data pool. The purpose of optimization can be to improve the quality of the sample set, increase the diversity of samples, or improve the training efficiency of the model. In the e-commerce advertising scenario, the server can use the preprocessed main sample instance set to optimize the original sample instance data pool. Specifically, the server can check whether the new preprocessed sample instance is repeated or similar to the existing samples in the data pool. If there are repeated or similar samples, the server can choose to retain samples with higher quality or more representativeness, while deleting or replacing samples with poor quality or redundant samples. In this way, the optimized sample instance data pool will contain more diverse and high-quality primary sample instances, which will help improve the accuracy and efficiency of advertising user analysis. Through the above steps, the server can continuously optimize and update the sample instance data pool using the preprocessed primary sample instance set, thereby ensuring the accuracy and effectiveness of advertising user analysis.

[0110] In the embodiment of the present invention, the sample instance set includes a main sample instance set and an auxiliary sample instance set, and the main sample instance set is stored in a sample instance data pool; the embodiment of the present invention also provides the following implementation manner.

[0111] When the preprocessed auxiliary sample instance set is determined, the preprocessed auxiliary sample instance set is saved to the sample instance data pool, and the sample instances in the preprocessed auxiliary sample instance set are added to the main sample instance set.

[0112] In an embodiment of the present invention, exemplarily, the sample instance set maintained by the server is divided into two parts: a main sample instance set and an auxiliary sample instance set. The main sample instance set is composed of sample instances that are critical to advertising user analysis. These sample instances are usually highly representative and important, and can directly reflect the user's advertising behavior characteristics. The auxiliary sample instance set contains some additional sample instances that may not be so critical but can also provide useful information. They can be used to enhance the diversity and richness of the main sample instance set. For example, in an e-commerce advertising scenario, the main sample instance set can contain a large number of key information such as user purchase history, browsing behavior, search keywords, etc., which can directly reflect the user's shopping preferences and interests. The auxiliary sample instance set can include user comments, shared content, likes, etc. on social media. Although this information is not as direct as purchase records, it can also provide valuable references for advertising user analysis. When the server collects new auxiliary sample instances, it will perform preprocessing operations on these instances. The purpose of preprocessing is mainly to clean data, extract features, convert formats, etc., for subsequent analysis and model training. The preprocessing operation can include steps such as removing duplicate data, filling missing values, removing noise data, text segmentation, and feature encoding. Taking the e-commerce advertising scenario as an example, after collecting the user's comment data on social media, the server will first clean the text of these comments to remove meaningless tags, special characters, etc. Then, the server will use natural language processing technology to segment and tag the comments, and extract the keywords and key phrases as features. Then, the server will encode and transform these features to make them suitable for the subsequent advertising user analysis model. After determining the pre-processed auxiliary sample instance set, the server will save these instances to the sample instance data pool and add them to the main sample instance set. The purpose of this is to expand the scale of the main sample instance set, increase the diversity and richness of the samples, and thus improve the accuracy and reliability of advertising user analysis. In the e-commerce advertising scenario, after pre-processing the user's comment data on social media, the server will save these data to the sample instance data pool. Then, the server will add these data to the main sample instance set, so that the main sample instance set not only contains direct information such as the user's purchase history, but also indirect information such as the user's behavior data on social media. In this way, the advertising user analysis model can more comprehensively understand the interests and needs of users, thereby providing a more accurate advertising push strategy. Through the above steps, the server can make full use of the auxiliary sample instance set to enhance the diversity and richness of the main sample instance set, and improve the accuracy and reliability of advertising user analysis. At the same time, by continuously collecting new auxiliary sample instances and performing preprocessing and adding operations, the server can also continuously update and optimize the sample instance set to adapt to the ever-changing user needs and advertising market environment.

[0113] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the above-mentioned advertising user analysis method. Figure 2 As shown, Figure 2 The block diagram of the computer device 100 provided in the embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112 and a communication unit 113. To achieve data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0114] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. An advertising user analysis method, characterized in that: The method comprises: In response to an advertising promotion operation triggered by a current user, obtaining a current user behavior of the current user in a preset monitoring period; Transmitting the current user behavior to the trained advertising user behavior recognition model, and obtaining a classification result of the current user behavior through the trained advertising user behavior recognition model; According to the classification result, determining the advertisement push strategy corresponding to the current user from a pre-stored advertisement push strategy library; Using the classification result and the advertisement push strategy as the advertisement user analysis result of the current user; The advertising user behavior recognition model is obtained by: Acquire a target behavior instance and at least two benchmark behavior categories, wherein the target behavior instance corresponds to at least two sample behavior categories; Obtain a baseline behavior category vector for each of the baseline behavior categories through a behavior feature extraction model, perform aggregation operations on at least two of the baseline behavior categories according to the baseline behavior category vectors to obtain at least two baseline behavior category centers, and obtain an aggregated behavior category for each of the baseline behavior category centers through the sample behavior categories; Transmitting the target behavior instance to an advertising user behavior recognition model, wherein the advertising user behavior recognition model includes a feature extraction component and a feature recognition component; Performing time-series segmentation processing on the target behavior instance to obtain a user behavior set, wherein the user behavior set includes at least two user interaction operations; Configuring an initial tag for an initial interaction operation of the user behavior set, and configuring a termination tag for a termination interaction operation of the user behavior set, to obtain a to-be-processed behavior set; Performing feature extraction processing on the set of behaviors to be processed to obtain a set of behavior operation vectors; By using a recurrent neural network, the feature extraction component performs a feature extraction operation on the behavior operation vector set to obtain an interaction feature vector set, wherein the interaction feature vector set includes a feature representation of each user interaction operation in the behavior set to be processed; By using a recurrent neural network, each of the feature representations is subjected to feature merging to obtain an instance feature vector; Transmitting the instance feature vector to the feature recognition component to obtain a first category reliability coefficient of each of the benchmark behavior category centers; Performing a standardization operation on each of the first category reliability coefficients to obtain the first category belonging confidence of each of the reference behavior category centers; In the case where the first category attribution confidence is not lower than a preset first confidence threshold, the benchmark behavior category center corresponding to the first category attribution confidence is used as the inferred behavior category center, and each of the benchmark behavior categories in the inferred behavior category center is traversed to calculate the vector distance between the instance feature vector and the benchmark behavior category vector of the benchmark behavior category; Performing a normalization operation on the vector distance to obtain a second category belonging confidence of the benchmark behavior category; In a case where the second category attribution confidence is not lower than a preset second confidence threshold, taking the benchmark behavior category corresponding to the second category attribution confidence as the benchmark behavior category corresponding to the target behavior instance; According to the aggregated behavior category of each of the benchmark behavior category centers, obtaining the aggregated behavior confidence of each of the benchmark behavior category centers; Calculating the mean square error deviation between each of the aggregated behavior confidences and the corresponding first category attribution confidence to obtain at least two confidence deviations; The sum of all the confidence deviations is taken as the first deviation; According to the sample behavior category, obtaining the target category confidence of each of the benchmark behavior categories in the inferred behavior category center; Calculating the mean square error deviation between each of the target category confidences and the corresponding second category belonging confidences to obtain at least two category deviations; The sum of all the category deviations is taken as the second deviation; According to the first deviation and the second deviation, an integrated training process is performed on the behavior feature extraction model and the advertising user behavior recognition model to obtain a trained advertising user behavior recognition model.

2. The method according to claim 1, characterized in that The target behavior instance is obtained by: Acquire a set of sample instances for training an advertisement user behavior recognition model, and determine a set of candidate instances to be screened based on the set of sample instances; Acquire a multi-task integrated model for performing instance screening processing on the candidate instance set; the multi-task integrated model is a model composed of multiple sub-task networks matching multiple evaluation indicators, and one evaluation indicator corresponds to one sub-task network; The candidate instance set is loaded into each subtask network of the multiple subtask networks, and each subtask network performs instance screening processing on the candidate instance set to obtain an instance screening result of the candidate instance set under each evaluation indicator; Preprocessing the sample instance set based on the instance screening results of the candidate instance set under each evaluation indicator; The target behavior instance is determined based on the preprocessed sample instance set.

3. The method according to claim 2, characterized in that The sample instance set includes a main sample instance set; the main sample instance set includes a plurality of sample instance data groups, each of the plurality of sample instance data groups includes at least one sample instance; The step of determining a set of candidate instances to be screened based on the set of sample instances includes: Get the number of samples used for instance sampling; Extracting sample instances corresponding to the sampling number from each of the sample instance data groups in sequence to obtain sample instance data groups corresponding to each of the sample instance data groups; Determine, based on the sampled sample instance data groups corresponding to the sampled sample instance data groups, the sampled sample instances that match the main sample instance set; The sampled sample instance is used as a sample instance in a candidate instance set, and the candidate instance set to be screened is determined based on the sample instance in the candidate instance set.

4. The method according to claim 2, characterized in that: The sample instance set includes a main sample instance set; the multiple evaluation indicators include a target evaluation indicator; the instance screening result under the target evaluation indicator includes an instance screening sub-result of each sample instance in the candidate instance set under the target evaluation indicator; The preprocessing of the sample instance set based on the instance screening result of the candidate instance set under each evaluation indicator includes: In the candidate instance set, the sample instances whose instance screening sub-results under the target evaluation index are abnormal are taken as abnormal sample instances; When the number of abnormal sample instances reaches a preprocessing condition, the target evaluation index is used as a pending evaluation index that needs to be preprocessed; Preprocessing is performed on the sample instances in the main sample instance set based on the evaluation index to be processed.

5. The method according to claim 2, characterized in that: The sample instance set includes an auxiliary sample instance set; the multiple evaluation indicators include a target evaluation indicator; the instance screening result under the target evaluation indicator includes an instance screening sub-result of each sample instance in the candidate instance set under the target evaluation indicator; The preprocessing of the sample instance set based on the instance screening result of the candidate instance set under each evaluation indicator includes: In the candidate instance set, the sample instances whose instance screening sub-results under the target evaluation index are abnormal are taken as abnormal sample instances; The abnormal sample instance is preprocessed based on the target evaluation index.

6. The method according to claim 2, characterized in that The target evaluation indicator among the multiple evaluation indicators includes multiple evaluation sub-indicators; the subtask network matching the target evaluation indicator is composed of multiple subtask units matching the multiple evaluation sub-indicators, and one evaluation sub-indicator corresponds to one subtask unit; The step of loading the candidate instance set into each subtask network of the plurality of subtask networks, and having each subtask network perform instance screening processing on the candidate instance set to obtain an instance screening result of the candidate instance set under each evaluation indicator includes: The candidate instance set is loaded into each subtask unit of the plurality of subtask units, and each subtask unit performs instance screening processing on the candidate instance set to obtain an instance screening result of the candidate instance set under each evaluation sub-indicator; Based on the instance screening results under each of the evaluation sub-indicators, determining the instance screening results of the candidate instance set under the target evaluation indicator; Based on the instance screening result of the candidate instance set under the target evaluation indicator, the instance screening result of the candidate instance set under each evaluation indicator is determined.

7. The method according to claim 2, characterized in that The sample instance set includes a plurality of sample instance data groups classified by user groups; the plurality of sample instance data groups classified by user groups include a first user group classified sample instance data group and a second user group classified sample instance data group under the first user group classified sample instance data group; The method further comprises: Obtaining a user group association rule for associating the first user group classification sample instance data group with the second user group classification sample instance data group; the user group association rule is used to characterize the corresponding relationship between the first user group classification sample instance data group and the second user group classification sample instance data group; Based on the user group association rule, summarizing the user group analysis results of the instance screening results of the candidate instance set under each evaluation indicator to obtain user group analysis data matching the second user group classification sample instance data group; Based on the user group association rule, summarizing the user group analysis results of the instance screening results of the candidate instance set under each evaluation indicator to obtain user group analysis data matching the first user group classification sample instance data group; The method further comprises: Based on the instance screening results of the candidate instance set under each evaluation indicator, determining quantitative indicator information at multiple multidimensional data analysis levels; Displaying a visualization interaction interface; the visualization interaction interface includes visualization components corresponding to the quantitative indicator information at the multiple multi-dimensional data analysis levels.

8. The method according to claim 2, characterized in that: The multiple evaluation indicators include a target evaluation indicator, and the multiple subtask networks include a target subtask network that matches the target evaluation indicator; The method further comprises: Acquire a sample screening instance that matches the target evaluation indicator; the sample screening instance is associated with a screening target value, and the screening target value is used to characterize the actual screening result of the sample screening instance under the target evaluation indicator; Acquire an original subtask network that matches the target subtask network, load the sample screening instance into the original subtask network, perform instance screening processing on the sample screening instance through the original subtask network, and obtain an inferred screening result of the sample screening instance under the target evaluation index; Performing a tuning operation on the original subtask network based on the real screening result and the inferred screening result, and using the original subtask network after the tuning operation as the target subtask network; The sample instance set includes a main sample instance set, and the main sample instance set is stored in a sample instance data pool; The method further comprises: When the preprocessed main sample instance set is determined, the main sample instance set stored in the sample instance data pool is optimized by using the preprocessed main sample instance set; The sample instance set also includes an auxiliary sample instance set, and the main sample instance set is stored in a sample instance data pool; The method further comprises: When the preprocessed auxiliary sample instance set is determined, the preprocessed auxiliary sample instance set is saved to the sample instance data pool, and the sample instances in the preprocessed auxiliary sample instance set are added to the main sample instance set.

9. A server system, characterized in that: The method comprises a server, wherein the server is used to execute the method described in any one of claims 1 to 8.

Images