Click prediction method based on genetic algorithm for feature interaction and pseudo-label learning

Abstract

A click prediction method based on genetic algorithm for feature interaction and pseudo-label learning includes data preprocessing, feature interaction, prediction model building, and marketing campaign prediction. This invention constructs new features through feature interaction based on genetic algorithm principles and combines pseudo-labeling technology to jointly model unlabeled and labeled data, resulting in a final binary classifier learner for predicting user clicks. This method effectively utilizes the iterative generation of high-quality new features using genetic algorithm and, combined with pseudo-labeling learning, effectively leverages unlabeled data, significantly improving the accuracy and robustness of user click behavior prediction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence in internet marketing, and more specifically, to a click prediction method based on genetic algorithms for feature interaction and pseudo-label learning. Background Technology

[0002] Click-through rate (CTR) prediction is a crucial aspect of internet computational advertising; the accuracy of user prediction directly impacts a company's marketing revenue.

[0003] The advertising process typically involves three stages: ad impressions, user clicks, and user registration (or conversion). A major objective of CTR prediction is to improve the efficiency from ad impressions to user clicks, i.e., to increase the click-through rate. Since user clicks or non-clicks are generally defined as tags of user behavior.

[0004] Those skilled in the art will understand that traditional CTR prediction can be viewed as a supervised learning problem: that is, after knowing a series of features such as user click behavior, user clicks are recorded as 1 and user non-clicks are recorded as 0, thereby performing binary classification learning.

[0005] CTR prediction involves forecasting a defined binary classification problem, generally categorized as supervised learning. In practice, the amount of data delivered is usually greater than the amount of valid data retrieved. Due to issues such as data transmission channels and network latency, not all delivered ads are actually sent to user terminals, and even if a user engages in related activities, this information cannot be guaranteed to be returned 100% of the time.

[0006] In other words, only a portion of the data distributed in each instance is ultimately labeled, i.e., valid. During model training, generally only these valid data are used. In fact, unlabeled data also contains a wealth of feature information related to user profiles, and it is easily obtained; furthermore, unlabeled data can improve the accuracy of classification decision boundaries, thereby enhancing the robustness of the model. Therefore, how to rationally and effectively utilize unlabeled data is a pressing issue that the industry urgently needs to address. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned defects in the existing technology and provide a click prediction method based on feature interaction and pseudo-label learning using genetic algorithms. This method constructs new features by interacting with features based on the idea of ​​genetic algorithms and combines pseudo-label technology to jointly model unlabeled and labeled data to obtain a binary classifier learner for predicting user clicks.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A click prediction method based on genetic algorithm for feature interaction and pseudo-label learning includes a data preprocessing step S1, a feature interaction step S2, and a prediction model building step S3.

[0010] The data preprocessing step S1 includes the following steps:

[0011] Step S11: Obtain the original information of M users, and extract the first original feature information from the original information of the users. The first original feature information includes Y features; wherein, the first original feature information includes user ID, user mobile phone number location, user access DPI, user access time and / or user click feature; wherein, the user mobile phone number location and user access DPI are category features, the user access time is a continuous feature, and, in the data collection process, both labeled data and unlabeled data are included, and M is greater than or equal to 2;

[0012] Step S12: Perform anomaly detection and processing on the user's original feature information;

[0013] Step S13: Process categorical features using one-hot encoding and continuous features using the RankGauss method;

[0014] Step S2, which generates the labeled training set, includes:

[0015] Based on the feature interaction of the first original feature information using a genetic algorithm, X high-quality second original feature information are selected; wherein, the number of features of the third original feature information is equal to the number of the first original feature information Y plus the number of the second original feature information X.

[0016] The prediction model establishment step S3 includes:

[0017] Step S31: Provide an initial neural network model; and separate the labeled data and unlabeled data with the third original feature information after sorting;

[0018] Step S32: Divide the labeled data into 80% as the first training set and 20% as the first validation set;

[0019] Step S33: Train the initial neural network model using the data from the first training set, and validate it using the first validation set to obtain the optimized intermediate neural network model;

[0020] Step S34: Based on the intermediate neural network model, predict each unlabeled data and output the probability of the predicted classification. If the probability of the predicted classification is greater than a certain threshold, the label of the unlabeled data is 1; otherwise, the label of the unlabeled data is 0. In this way, each unlabeled data is labeled with a predicted pseudo-label.

[0021] Step S35: 80% of the labeled data will be re-segmented as the second training set and 20% as the second validation set; the labeled data of the second training set will be mixed with the pseudo-labeled data obtained in step S34 to form a new training set.

[0022] Step S36: Train and validate the initial neural network model using each sample in the new training set to obtain a neural network model with optimized parameters, and then validate it using the validation set to obtain the final prediction network model.

[0023] Further, step S2 includes the following steps:

[0024] Step S21: Establish a binary tree based on the interaction operation between Y features. Randomly initialize N trees, and the depth of each tree is an adjustable parameter.

[0025] Step S22: Calculate the corresponding fitness function value for each generated tree using the feature interaction operation model;

[0026] Step S23: Determine whether the fitness function value satisfies the convergence condition. If convergence is satisfied, output the corresponding feature combination of the tree and execute step S26; if convergence is not satisfied, execute step S24.

[0027] Step S24: Perform selection, crossover, and mutation operations on the tree;

[0028] Step S25: Repeat steps S22 to S24 until convergence;

[0029] Step S26: Continue until the work of selecting high-quality interaction features for each tree is completed, resulting in X selected high-quality interaction features, thereby determining the number of features of the third original feature information as X+Y.

[0030] Furthermore, the fitness function of the genetic algorithm is the cross-entropy loss.

[0031] Furthermore, the click prediction method also includes a marketing campaign prediction step S4, which specifically includes:

[0032] Step S41: Obtain the user group for the marketing campaign click prediction and the original user information of the user group, and extract the first original feature information from the original user information. The first original feature information includes Y features. The first original feature information includes user ID, user mobile phone number location, user access DPI, user access time and / or user click feature. In addition, the data collection process includes both labeled data and unlabeled data.

[0033] Step S42: Perform anomaly detection and processing on the first original information of each user; process the categorical features through one-hot encoding and process the continuous features using the RankGauss method;

[0034] Step S43: Based on the feature interaction of the first original feature information using a genetic algorithm, select X high-quality second original feature information; wherein, the number of features of the third original feature information is equal to the number of the first original feature information Y plus the number of the second original feature information X;

[0035] Step S44: Provide a pre-trained prediction network model, and use the third original feature information in sequence to obtain the predicted values ​​of all users in the user group for the marketing campaign click prediction; wherein, the model predicted value of the user is the user's click intention degree.

[0036] Furthermore, the model prediction step S4 also includes:

[0037] Step S45: Based on actual campaign needs, select all or a portion of the users whose model prediction values ​​show a click intent of 1 for targeted marketing.

[0038] Furthermore, step S3 also includes step S37, which involves processing and optimizing the prediction network model using model evaluation metrics; the model evaluation metrics include AUC value, Log loss value, and relative information gain (RIG) value.

[0039] Furthermore, the model optimization process includes one or more of the following:

[0040] Add batch normalization to resolve the issue of internal covariate offset in the data;

[0041] Add a feature to the network that allows some neurons to remain dormant during training;

[0042] Adjusting the learning rate typically involves using strategies such as exponential decay to adjust the learning rate during the training process.

[0043] To improve the problem of insufficient generalization ability caused by large data variance, multiple sub-training methods are set up and averaged.

[0044] Adding L1 or L2 regularization penalizes the loss function to reduce the risk of overfitting;

[0045] Methods for optimizing hyperparameters.

[0046] As can be seen from the above technical solutions, the click prediction method based on genetic algorithm for feature interaction and pseudo-label learning of the present invention can effectively utilize the genetic algorithm to generate high-quality new features through iteration, and combined with pseudo-label learning, effectively utilize unlabeled data, significantly improving the accuracy and robustness of predicting user click behavior. Attached Figure Description

[0047] Figure 1 The image shows click prediction based on feature interaction and pseudo-label learning using a genetic algorithm in an embodiment of the present invention.

[0048] Flowchart of the method

[0049] Figure 2 The diagram shown is a binary tree diagram for interactive operations between features in an embodiment of the present invention.

[0050] Figure 3 The diagram shown is a flowchart of the genetic algorithm for feature interaction in an embodiment of the present invention.

[0051] Figure 4 The diagram shown illustrates the crossover operation of the feature tree in the genetic algorithm of this invention.

[0052] Figure 5 The diagram shown is a schematic of the pseudo-label learning framework in an embodiment of the present invention. Detailed Implementation

[0053] The following is in conjunction with the appendix Figure 1-5 The specific embodiments of the present invention will be further described in detail below.

[0054] In the following detailed description of the embodiments of the present invention, in order to clearly illustrate the structure of the present invention for ease of explanation, the structures in the accompanying drawings are not drawn to scale and have been partially enlarged, deformed and simplified. Therefore, it should be avoided to interpret this as a limitation of the present invention.

[0055] It should be noted that in the embodiments of this invention, for unlabeled data, we do not need to give its precise category, but only a fuzzy approximate label, which is beneficial for the neural network model to learn features effectively. Therefore, this invention, based on existing data and combined with unlabeled data, improves the overall performance of the model in CTR prediction, and has important theoretical and application value in practical business.

[0056] Please see Figure 1 , Figure 1 The diagram shows a flowchart of the click prediction method based on genetic algorithms for feature interaction and pseudo-label learning, as described in this invention. Figure 1 As shown, the prediction method includes a data preprocessing step S1, a labeled data training set generation step S2, a prediction model building step S3, and a marketing campaign prediction step S4.

[0057] In embodiments of the present invention, the data preprocessing step is very important, and the data preprocessing step S1 includes, for example:

[0058] Next steps:

[0059] Step S11: Obtain the original information of M users, and extract the first original feature information from the original information of the users. The first original feature information includes Y features; wherein, the first original feature information includes user ID, user mobile phone number location, user access DPI, user access time and / or user click feature; wherein, the user mobile phone number location and user access DPI are category features, the user access time is a continuous feature, and, in the data collection process, both labeled data and unlabeled data are included, and M is greater than or equal to 2.

[0060] Please refer to Table 1 below. Table 1 is a tabular description of the original data before preprocessing. The original data before preprocessing is shown in Table 1 below:

[0061] Table 1

[0062] User ID Access DPI Place of Origin Access time Click or not? 199990 dpi1 HN 150702957019 1 199991 dpi3 HN 150702957147 1 199992 dpi2 HN 150702960147 1 … … … … … 299991 dpi4 HN 150702953220 1 299992 dpi3 LN 150702158424 0 299993 dpi5 LN 150702983485 0 299994 dpi1 LN 150712478952 0 … … … … …

[0063] Please note that at this point, for the labels in the labeled data, i.e. whether the user clicked, they are only distinguished as positive samples (1) and unlabeled samples (0). The unlabeled data format (i.e., features) is the same as the labeled data format, except that it lacks labels.

[0064] In embodiments of the present invention, the aforementioned first original feature information also needs to undergo steps such as anomaly detection and processing, category feature processing, continuous feature processing, and dimensionality reduction processing.

[0065] Step S12: Perform anomaly detection and processing on the user's original information.

[0066] Anomaly Detection and Handling: In accordance with business requirements, missing values ​​and excessively large values ​​in the initial raw feature information need to be deleted or filled. During data collection, since the number of users is typically in the millions, missing values ​​may occur. If the amount of missing data is small, it can generally be removed directly. If it is unclear whether missing data will affect the final model training effect, missing values ​​can generally be filled by taking the mean, mode, or median.

[0067] Furthermore, during data collection, we may encounter issues with excessively large values, such as a user accessing [number] times within a single day.

[0068] In cases where the DPI reaches tens of thousands, this generally does not particularly help improve the generalization ability of the model in actual modeling. Therefore, it is also possible to use elimination or imputation methods to handle the situation accordingly.

[0069] As can be seen from the table above, compared to traditional user access data, each user ID corresponds to multiple DPI access records, and the access time is also given.

[0070] Step S13: Process other continuous or categorical features; process categorical features using one-hot encoding and continuous features using the RankGauss method.

[0071] Categorical Feature Processing: For categorical features such as user number location, one-hot encoding can be used for expansion. For example, if the location is (HN, LN, SH), after one-hot encoding, the location of a particular user will be 1, and other locations will be 0. Taking user ID 199990 as an example, its location is HN, so after one-hot encoding, its location vector is (1,0,0); correspondingly, the location vectors of users 199991 and 199992 are (0,1,0) and (0,0,1) respectively.

[0072] Continuous Feature Processing: For continuous features such as user access time, the RankGauss method can be used to adjust the data distribution. The RankGauss method is used for adjusting the data distribution and / or dimensionality reduction of continuous features; principal component analysis is used to reduce the dimensionality of high-dimensional features after categorical feature processing.

[0073] In embodiments of this invention, Principal Component Analysis (PCA) is used to reduce the dimensionality of high-dimensional features. As described above regarding the processing of categorical features, one-hot encoding typically results in a high-dimensional sparse matrix. For neural network training, this means that differentiation is impossible in many places during error backpropagation, which is clearly detrimental to network training. Furthermore, high-dimensional features increase computational overhead. Therefore, it is necessary to reduce the dimensionality of high-dimensional features first. PCA achieves dimensionality reduction by finding the maximum variance of the original data in a certain projection direction; while reducing feature dimensionality, it minimizes the loss of information contained in the first original features, thereby achieving the goal of comprehensive analysis of the collected data.

[0074] After the above data processing steps, feature interaction step S2 can be performed. In this embodiment of the invention, feature interaction is based on a genetic algorithm. Taking the pairwise interaction between two features Fet0 and Fet1 in the first original feature information as an example, the features can follow arithmetic operations (addition, subtraction, multiplication, division); logical operations (AND, OR); and exponential and polynomial operations.

[0075] Theoretically, some operational rules can generate a series of new features (second original feature information) for training, while some operational rules are meaningless for actual business and will bring additional noise to the neural network model. Therefore, in the embodiments of the present invention, the interaction features are iterated by rules based on genetic algorithms, following the natural law of "survival of the fittest", thereby selecting high-quality second original feature information.

[0076] Please see Figure 2 , Figure 2 The diagram shown is a binary tree diagram illustrating the interactive operations between features in an embodiment of the present invention. All operations performed between the first original features are described as follows: Figure 2 The example shown is a binary tree. Figure 2 In the binary tree diagram shown, all leaf nodes (i.e., the last level) represent input variables or constants, while the internal nodes represent the operations performed by feature interactions.

[0077] The core of genetic algorithms lies in measuring the fitness after feature interaction operations. The fitness function is similar to the objective function or loss function in machine learning algorithms. Preferably, for binary classification problems involving CTR prediction, cross-entropy loss can be used as the fitness function of the genetic algorithm.

[0078] Please see Figure 3 , Figure 3 The diagram shown is a flowchart of the genetic algorithm for feature interaction in an embodiment of the present invention. Figure 3 As shown, step S2 may specifically include the following steps:

[0079] Step S21: Establish a binary tree based on the interaction operation between Y features. Randomly initialize N trees, and the depth of each tree is an adjustable parameter.

[0080] Step S22: Calculate the corresponding fitness function value for each generated tree using the feature interaction operation model;

[0081] Step S23: Determine whether the fitness function value satisfies the convergence condition. If convergence is satisfied, output the corresponding feature combination of the tree and execute step S26; if convergence is not satisfied, execute step S24.

[0082] Step S24: Perform selection, crossover, and mutation operations on the tree;

[0083] Step S25: Repeat steps S22 to S24 until convergence;

[0084] Step S26: Continue until the work of selecting high-quality interaction features for each tree is completed, resulting in X selected high-quality interaction features, thereby determining the number of features of the third original feature information as X+Y.

[0085] In the genetic algorithm described above, crossover, mutation, and mutation are the key steps in the "evolution" of features. Here, we will focus on the crossover operation. Crossover of a feature binary tree refers to randomly selecting a subtree and replacing it with a random subtree of another feature tree. This other feature tree is usually the one with the highest fitness among the remaining feature trees.

[0086] Please see Figure 4 , Figure 4 The diagram shown illustrates the crossover operation of the feature tree in the genetic algorithm of this invention. Figure 4 As shown, the mutation operation of a feature tree refers to the complete replacement of its subtrees with another randomly initialized tree. Through continuous evolution using a genetic algorithm, high-quality interaction features can be selected, facilitating the next step of model training.

[0087] After the above generation steps are completed, the prediction model building step S3 can be executed. Specifically, this includes training and validating the neural network model with each sample in the new training set to obtain the neural network model with optimized parameters, and then validating it with a validation set to obtain the final prediction network model.

[0088] Specifically, please refer to Figure 5 , Figure 5 The diagram shown is a schematic representation of the pseudo-label learning framework in an embodiment of the present invention. Figure 5 As shown, the prediction model establishment step S3 includes:

[0089] Step S31: Provide an initial neural network model; and separate the labeled data and unlabeled data with the third original feature information after sorting;

[0090] Step S32: Divide the labeled data into 80% as the first training set and 20% as the first validation set;

[0091] Step S33: Train the initial neural network model using the data from the first training set, and validate it using the first validation set to obtain the optimized intermediate neural network model;

[0092] Step S34: Based on the intermediate neural network model, predict each unlabeled data and output the probability of the predicted classification. If the probability of the predicted classification is greater than a certain threshold, the label of the unlabeled data is 1; otherwise, the label of the unlabeled data is 0. In this way, each unlabeled data is labeled with a predicted pseudo-label.

[0093] Step S35: 80% of the labeled data will be re-segmented as the second training set and 20% as the second validation set; the labeled data of the second training set will be mixed with the pseudo-labeled data obtained in step S34 to form a new training set.

[0094] Step S36: Train and validate the initial neural network model using each sample in the new training set to obtain a neural network model with optimized parameters, and then validate it using the validation set to obtain the final prediction network model.

[0095] In an embodiment of the present invention, step S3 may further include performing model evaluation index processing and optimization processing on the marketing activity prediction model based on the click prediction method model using genetic algorithm for feature interaction and pseudo-label learning (step S37); the model evaluation index includes AUC value, Log loss value and relative information gain (RIG) value.

[0096] The model evaluation metrics include AUC (Area Under Curve) value, Log loss value, and relative information gain.

[0097] RIG (Relative Information Gain) value. Generally, the closer the AUC value is to 1, the better the model's classification performance. The smaller the Log loss value, the higher the accuracy of click-through rate prediction; the larger the relative information gain value, the better the model's performance.

[0098] For example, after processing the data according to the above steps and training the model, the AUC value obtained through local validation can be used to determine the training effect of the model. If the effect is poor, the model generally needs to be tuned. For deep learning algorithms, optimization can generally be carried out in the following aspects:

[0099] ① Add batch normalization to solve the problem of internal covariate shift in the data.

[0100] ② Add Dropout to the network, which allows some neurons to remain dormant during training.

[0101] ③ Adjusting the learning rate: This is usually done through strategies such as exponential decay during the training process.

[0102] ④ Set up multiple sub-training methods and take the average to reduce the risk of overfitting during the training process.

[0103] ⑤ Add L1 or L2 regularization to penalize the loss function and reduce the risk of overfitting.

[0104] ⑥ Hyperparameter optimization.

[0105] In optimizing hyperparameters, grid search or random search are commonly used; however, these methods are computationally expensive and inefficient. In this embodiment, Bayesian optimization is employed. Bayesian optimization calculates the posterior probability distribution of the first n data points using Gaussian process regression, obtaining the mean and variance of each hyperparameter at each value point. By balancing the mean and variance, and based on the joint probability distribution among the hyperparameters, Bayesian optimization ultimately selects a better set of hyperparameters.

[0106] After all the above processing steps are completed, the features can be fed into the trained user prediction model to screen out a subset of users with high interest before advertising is launched, and to accurately target these users with marketing ads. That is, the present invention may also include a marketing activity prediction step S4, which specifically includes:

[0107] Step S41: Obtain the user group for the marketing campaign click prediction and the original user information of the user group. The first original feature information includes Y features. The first original feature information includes user ID, user mobile phone number location, user access DPI, user access time and / or user click feature. In addition, the data collection process includes both labeled data and unlabeled data.

[0108] Step S42: Perform anomaly detection and processing on the first original information of each user; process the categorical features through one-hot encoding and process the continuous features using the RankGauss method;

[0109] Step S43: Based on the feature interaction of the first original feature information using a genetic algorithm, select X high-quality second original feature information; wherein, the number of features of the third original feature information is equal to the number of the first original feature information Y plus the number of the second original feature information X;

[0110] Step S44: Provide a pre-trained prediction network model, and use the third original feature information sequentially to obtain the predicted values ​​of all users in the user group for the marketing campaign click prediction; wherein, the model prediction value of each user is the user's click intention degree;

[0111] Step S45: Based on actual campaign needs, select all or a portion of the users whose model prediction values ​​show a click intent of 1 for targeted marketing.

[0112] The results show that the method of this invention can effectively utilize WaveNet to generate new feature sequences in the case of small samples, while the simulated annealing algorithm ensures the effectiveness of the new sequences, significantly improving the accuracy of predicting user click behavior.

[0113] The above description is merely a preferred embodiment of the present invention. The embodiments are not intended to limit the scope of patent protection of the present invention. Therefore, any equivalent structural changes made based on the description and drawings of the present invention should also be included within the scope of protection of the present invention.

Claims

1. A click prediction method based on genetic algorithm for feature interaction and pseudo-label learning, characterized in that, It includes data preprocessing step S1, feature interaction step S2, and prediction model building step S3; The data preprocessing step S1 includes the following steps: Step S11: Obtain the original information of M users, and extract the first original feature information from the original information of the users. The first original feature information includes Y features. The first original feature information includes user ID, user mobile phone number location, user access DPI, user access time and / or whether the user clicked. In addition, the data collection process includes both labeled data and unlabeled data, and M is greater than or equal to 2. Step S12: Perform anomaly detection and processing on the user's original initial feature information; Step S13: Process categorical features using one-hot encoding and continuous features using the RankGauss method; Step S2, which generates the labeled training set, includes: Based on the feature interaction of the first original feature information using a genetic algorithm, X high-quality second original feature information are selected; wherein, the number of features of the third original feature information is equal to the number of the first original feature information Y plus the number of the second original feature information X. The prediction model establishment step S3 includes: Step S31: Provide an initial neural network model; and separate the labeled data and unlabeled data with the third original feature information after sorting; Step S32: Divide the labeled data into 80% as the first training set and 20% as the first validation set; Step S33: Train the initial neural network model using the data from the first training set, and validate it using the first validation set to obtain the optimized intermediate neural network model; Step S34: Based on the intermediate neural network model, predict each unlabeled data and output the probability of the predicted classification. If the probability of the predicted classification is greater than a certain threshold, the label of the unlabeled data is 1; otherwise, the label of the unlabeled data is 0. In this way, each unlabeled data is labeled with a predicted pseudo-label. Step S35: 80% of the labeled data will be re-segmented as the second training set and 20% as the second validation set; the labeled data of the second training set will be mixed with the pseudo-labeled data obtained in step S34 to form a new training set. Step S36: Train and validate the initial neural network model using each sample in the new training set to obtain a neural network model with optimized parameters, and then validate it using the validation set to obtain the final prediction network model.

2. The click prediction method according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Establish a binary tree based on the interaction operation between Y features. Randomly initialize N trees, and the depth of each tree is an adjustable parameter. Step S22: Calculate the corresponding fitness function value for each generated tree using the feature interaction operation model; Step S23: Determine whether the fitness function value satisfies the convergence condition. If convergence is satisfied, output the corresponding feature combination of the tree and execute step S26; if convergence is not satisfied, execute step S24. Step S24: Perform selection, crossover, and mutation operations on the tree; Step S25: Repeat steps S22 to S24 until convergence; Step S26: Continue until the work of selecting high-quality interaction features for each tree is completed, resulting in X selected high-quality interaction features, thereby determining the number of features of the third original feature information as X+Y.

3. The click prediction method according to claim 2, characterized in that, The fitness function of the genetic algorithm is cross-entropy loss.

4. The click prediction method according to claim 1, characterized in that, It also includes marketing campaign forecasting step S4, which specifically includes: Step S41: Obtain the user group for the marketing campaign click prediction and the original user information of the user group. The first original feature information includes Y features. The first original feature information includes user ID, user mobile phone number location, user access DPI, user access time and / or user click feature. In addition, the data collection process includes both labeled data and unlabeled data. Step S42: Perform anomaly detection and processing on the first original information of each user; process the categorical features through one-hot encoding and process the continuous features using the RankGauss method; Step S43: Based on the feature interaction of the first original feature information using a genetic algorithm, select X high-quality second original feature information; wherein, the number of features of the third original feature information is equal to the number of the first original feature information Y plus the number of the second original feature information X; Step S44: Provide a pre-trained prediction network model, and use the third original feature information in sequence to obtain the predicted values ​​of all users in the user group for the marketing campaign click prediction; wherein, the model predicted value of the user is the user's click intention degree.

5. The click prediction method according to claim 4, characterized in that, The model prediction step S4 further includes: Step S45: Based on actual campaign needs, select all or a portion of the users whose model prediction values ​​show a click intent of 1 for targeted marketing.

6. The click prediction method according to claim 1, characterized in that, Step S3 further includes step S37, which involves processing and optimizing the prediction network model using model evaluation metrics; the model evaluation metrics include AUC value, Logloss value, and relative information gain (RIG) value.

7. The click prediction method according to claim 6, characterized in that, The model optimization process includes one or more of the following: Add batch normalization to resolve the issue of internal covariate offset in the data; Add a feature to the network that allows some neurons to remain dormant during training; Adjusting the learning rate typically involves using strategies such as exponential decay to adjust the learning rate during the training process. To improve the problem of insufficient generalization ability caused by large data variance, multiple sub-training methods are set up and averaged. Adding L1 or L2 regularization penalizes the loss function to reduce the risk of overfitting; Methods for optimizing hyperparameters.

Images