Predictive ensemble modeling method based on unlabeled sample learning

Abstract

A predictive ensemble modeling method based on unlabeled sample learning includes a data preprocessing step, a feature engineering processing step, a positive / unlabeled sample random segmentation and combined ensemble training strategy step, and a model training step; wherein the present invention randomly samples positive label samples and unlabeled samples in the original training data and combines them into k new sample spaces according to a certain sample ratio; at the same time, a portion of the data is retained as out-of-bag data for verification; k learners are used for training and to establish k prediction models; for each of the established k prediction models, independent out-of-bag data is predicted, and each sample in the out-of-bag data can obtain k prediction results; the k prediction results are averaged as the final prediction result of the out-of-bag data; the k prediction models are subjected to model evaluation index processing and tuning processing to obtain the final k prediction models. Therefore, the present invention effectively improves the effect of feature screening.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence in Internet marketing, and more specifically, to a predictive ensemble modeling method based on unlabeled sample learning. Background Art

[0002] Click-Through Rate (CTR) prediction is a critical issue in internet computing advertising; the accuracy of user predictions directly impacts a company's marketing advertising revenue. Click-Through Rate (CTR) prediction is a critical issue in internet computing advertising; the accuracy of user predictions directly impacts a company's marketing advertising revenue.

[0003] The advertising process generally involves three steps: ad exposure, user clicks, and user registration (or conversion). A major goal of CTR estimation is to improve the efficiency from ad exposure to user clicks, that is, to increase click-through rate (CTR). User clicks or non-clicks are generally defined as labels for user behavior.

[0004] Those skilled in the art are aware that traditional CTR estimation 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] When estimating click-through rates (CTRs), sample data often suffers from a significant distribution imbalance, meaning that positive samples (users who click) typically only account for a small fraction of the total sample. Generally speaking, users who don't click are assigned a negative label, creating a typical binary classification problem along with the positive samples. However, based on actual advertising experience, users who don't click don't necessarily have no intention of clicking, so simply classifying them as negative samples is inaccurate.

[0006] Compared to binary classification problems with clear labels of clicks and no clicks, this problem uses a dataset of positive and unlabeled samples, making it possible to use semi-supervised learning methods. However, due to the large proportion of unlabeled samples, directly applying model training often yields poor results. Summary of the Invention

[0007] The purpose of the present invention is to provide a predictive ensemble modeling method based on unlabeled sample learning, and to develop a learning method suitable for unlabeled samples in click-through rate estimation, which can effectively avoid the above risks and improve the user click-through rate prediction results.

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

[0009] A predictive ensemble modeling method based on unlabeled sample learning includes a data preprocessing step S1, a feature engineering processing step S2, a positive / unlabeled sample random segmentation and combined ensemble training strategy step S3, and a model training step S4;

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

[0011] Step S11: Obtaining raw data information of N users, and extracting raw feature information from the raw data information of the users; wherein the raw feature information includes user ID, user province, user city, user access DPI, user access time (frequency), device type, and / or whether the user clicks; wherein the user province, user city, and user access DPI are categorical features, the user access time is a continuous feature, and N is greater than or equal to 2;

[0012] Step S12: performing abnormality detection and processing on the original feature information of the user;

[0013] Step S13: Process the categorical features through one-hot encoding, perform normalization and gradient descent processing according to business needs and solution model selection, or use the RankGauss method to adjust the sample distribution of continuous features;

[0014] The feature engineering processing step S2 includes the following steps:

[0015] Step S21: User access DPI processing step, creating a feature column of user ID and user access DPI;

[0016] Step S22: Processing the category features by one-hot encoding;

[0017] Step S23: Perform numerical feature transformation on continuous features;

[0018] Step S24: performing feature screening and feature construction to form new features;

[0019] Step S25: performing continuous feature discretization processing;

[0020] Step S3 specifically includes the following steps:

[0021] Randomly sample the positive label samples and unlabeled samples in the original training data and combine them into k new sample spaces according to a certain sample ratio; at the same time, retain a part of the data as out-of-bag data for verification;

[0022] The model training step S4 includes:

[0023] Step S41: treat the unlabeled samples in the k pieces of data as negative samples, and form a new training data set with the positive samples, and use the out-of-bag data as the validation data set;

[0024] Step S42: Using k learners to train and establish k prediction models based on the training data set and the validation data set;

[0025] Step S43: For each of the k established prediction models, prediction is performed on the independent out-of-bag data in step S41, and each sample in the out-of-bag data can obtain k prediction results; wherein, the k prediction results are a probability value;

[0026] Step S44: averaging the k prediction results as the final prediction result of the out-of-bag data;

[0027] Step S45: Based on the final prediction results of the out-of-bag data, the k prediction models are subjected to model evaluation index processing and optimization processing to obtain the final k prediction models.

[0028] Furthermore, the ratio of the training set samples to the validation set samples is adjusted according to the validation effect of the model validation set samples.

[0029] Furthermore, the k prediction models are gradient boosting decision tree models, and their implementation methods include XGBoost or lightGBM.

[0030] Furthermore, the model evaluation indicators include AUC value, Log loss value and relative information gain RIG value.

[0031] Furthermore, the prediction method further includes a prediction step S5, which specifically includes:

[0032] Step S51: Obtain a user group for which click prediction for a marketing campaign is to be performed and a sample of users of the user group to be predicted, and extract original feature information from the sample to be predicted; wherein the original feature information includes user ID, user province, user city, user access DPI, user access time, device type, and / or whether the user clicks; wherein the user province, user city, and user access DPI are categorical features, the user access time is a continuous feature, and N is greater than or equal to 2;

[0033] Step S52: Perform anomaly detection and processing on the raw data information of each user; process the categorical features through one-hot encoding, perform normalization and gradient descent processing according to business needs and solution model selection, or use the RankGauss method to adjust the sample distribution of continuous features;

[0034] Step S53: Provide the final k trained prediction models, input the characteristics of each user to be evaluated into the k estimation models in turn, obtain k prediction results, and finally take the average of these k prediction results to obtain the prediction values of all the users in the user group for click prediction of the marketing activity; wherein the model prediction value of the user is the click intention of the user.

[0035] Furthermore, the model prediction step S5 further includes:

[0036] Step S54: According to actual delivery needs, select all or part of the users whose model predicted values have a concentrated click intention of 1 to perform precision marketing tasks.

[0037] It can be seen from the above technical solution that the integrated modeling method for advertising click-through rate prediction based on unlabeled sample learning of the present invention randomly divides and combines positive label samples and unlabeled samples in the data set to form a multi-fold training sample space, and then predicts an independent out-of-bag sample after learner training. It integrates the full learning of different learners on the characteristics of different sample spaces, effectively improves the effect of feature screening, has important practical significance for improving the prediction effect, and has obtained improved click-through rate prediction results in actual business. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The figure shows a flow chart of an integrated modeling method for predicting click-through rate of advertisements based on unlabeled sample learning in an embodiment of the present invention.

[0039] Figure 2 The figure shows a process diagram of an unsupervised learning strategy based on random sampling and ensemble learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following is combined with Figure 1-2 , the specific implementation methods of the present invention are further described in detail.

[0041] In the following specific embodiments, when describing the embodiments of the present invention in detail, in order to clearly represent the structure of the present invention for the convenience of explanation, the structures in the accompanying drawings are not drawn according to general proportions, and are partially enlarged, deformed and simplified. Therefore, it should be avoided to understand this as a limitation of the present invention.

[0042] See also Figure 1 , Figure 1 The figure shows a flow chart of the integrated modeling method for predicting click-through rate of advertisements based on unlabeled sample learning. Figure 1As shown, it includes data preprocessing step S1, feature engineering processing step S2, positive / unlabeled sample random segmentation and combination integration training strategy step S3, and model training and model establishment step S4.

[0043] In the embodiment of the present invention, the data preprocessing step S1 is very important, and the data preprocessing step S1 includes the following steps:

[0044] Step S11: Obtain the original data information of N users, and extract original feature information from the original data information of the users; wherein the original feature information includes user ID (user_id), user's province (province), user's city (city), user's access DPI (dpi), user's access time (frequency), device type (phone_type) and / or whether the user clicks on the feature; wherein the user's province, user's city, and user's access DPI are category features, the user's access time is a continuous feature, and N is greater than or equal to 2.

[0045] Please refer to Table 1 below. Table 1 is a table description of the raw data information before preprocessing. The format of the raw data information before preprocessing is shown in Table 1 below:

[0046] Table 1

[0047]

[0048] Please note that at this time, for the label of the data, that is, whether the user clicks, only positive samples (1) and unlabeled samples (0) are distinguished.

[0049] In an embodiment of the present invention, the above-mentioned raw data information also needs to go through steps such as anomaly detection and processing, category feature processing, continuous feature processing and dimensionality reduction processing.

[0050] Step S12: performing abnormality detection and processing on the original data information of the user.

[0051] Anomaly detection and processing: In the process of integrating business requirements, missing values and oversized values in the raw data need to be deleted or filled. During the raw data collection process, due to the average user volume in the millions, missing values may occur. If the missing value is small, it can generally be directly eliminated. If it is difficult to determine whether the missing raw data will affect the final model training results, the missing values can generally be filled by taking the mean, mode, and median.

[0052] In addition, during the process of collecting raw data information, you may encounter problems with excessively large values. For example, a user visits the DPI tens of thousands of times in a day. This situation is generally not particularly helpful in improving the generalization ability of the model during the actual modeling process. Therefore, you can also use elimination or filling methods to perform corresponding processing.

[0053] Step S13: Process the categorical features through one-hot encoding, perform normalization and gradient descent processing according to business needs and solution model selection, or use the RankGauss method to adjust the sample distribution of continuous features.

[0054] As can be seen from the above table, compared with traditional user access data, each user ID corresponds to multiple DPI access records, and the access time is also given.

[0055] This embodiment of the present invention also includes processing user-accessed DPIs. Because each user accesses a unique DPI, the same user ID may appear multiple times in the original data, indicating that the user has multiple access records. Specifically, after deduplicating all DPIs, a feature column for user-accessed DPIs is created. If a user accesses several DPIs, the statistical count for those DPIs is incremented by one.

[0056] Step S2: Feature engineering processing step, which performs feature screening and feature construction on the user's original data information to form new features; it specifically includes the following steps:

[0057] Step S21: User access DPI processing step, creating a feature column of user ID and user access DPI. Because each user's accessed DPI is not unique, the same user ID may appear multiple times in the original data information, meaning that the user has multiple access records. Therefore, after deduplicating all DPIs, a feature column of user DPI is created. If a user accesses several DPIs, the statistical count of these DPIs is incremented by 1.

[0058] Step S22: Categorical feature processing: Categorical features such as the user's province, city, and device type can be expanded using one-hot encoding. For example, the provinces are (HN, JS, LN, SH, JL, BJ). After one-hot encoding, the province of a particular user is 1, while the others are 0. For example, user ID 199990, whose province is HN, has a one-hot encoding vector of (1, 0, 0, 0, 0). Correspondingly, the province vector of user 199991 is (0, 1, 0, 0, 0).

[0059] Step S23: Perform numerical feature transformation on continuous features.

[0060] This step, called numerical feature transformation, can be tailored to business needs and model selection. For example, for models using gradient descent, such as logistic regression and neural networks, normalization generally results in smoother gradient descent and faster convergence. For another example, some machine learning models assume that features follow a Gaussian distribution; in this case, methods like GaussRank can be used to adjust the sample distribution.

[0061] Step S24: Perform feature screening and feature construction to form new features.

[0062] Feature screening: Based on the embedding method, a random forest-based learner is established to train the preprocessed data and output the importance of each feature. Several features with lower weights are eliminated, and then the next round of training is carried out based on the new feature set. The recursive elimination of features with lower importance is repeated M times. Preferably, M is 3.

[0063] Feature construction: Add commonly used statistical features such as skewness, peakedness, etc.

[0064] Step S25: Discretize continuous features.

[0065] For some machine learning algorithms, such as logistic regression and feedforward neural networks, discretizing continuous features based on multiple thresholds can effectively increase the convergence speed of algorithm optimization, prevent the impact of outliers on model training, and, to a certain extent, play a regularization role. For example, based on the different quantiles of the sample distribution on a continuous feature (assuming the quantiles are 0.25, 0.5, and 0.75), the samples can be divided into five intervals on the continuous feature. Each sample may be distributed in one of the five intervals, and the sample takes the value of 1 in the feature interval it is located in and the value of 0 in the other four intervals. The effect of discretizing a continuous feature is to expand a continuous feature into a one-hot encoding form.

[0066] After preprocessing and feature engineering, the data is available as shown in Table 2 below:

[0067] Table 2

[0068]

[0069]

[0070] Among them, continuous features include statistical features such as DPI access frequency, kurtosis, and skewness. Categorical features include the one-hot encoding of the user's DPI access, the one-hot encoding of the number's location, and the discretized DPI access frequency.

[0071] Please combine Figure 1 See Figure 2 The prediction model establishment step S3 includes the steps of random segmentation of positive / unlabeled samples and combined integrated training strategy.

[0072] In the embodiment of the present invention, in view of the uneven distribution of positive label samples and unlabeled samples in click-through rate estimation, the original supervised learning strategy is changed and an unsupervised learning strategy based on random sampling and ensemble learning is established.

[0073] Specifically, this unsupervised learning strategy treats users who clicked on ads in the original dataset as positively labeled sample data, while users who did not click on ads are treated as unlabeled sample data (unlabeled means it is not confirmed whether the user will click). The positively labeled and unlabeled samples in the original training data are randomly sampled and combined into k new sample spaces at a certain sample ratio.

[0074] At the same time, a portion of the data is retained as out-of-bag data (oob data) for validation. The ratio of training set samples to validation set samples can be adjusted based on the validation results of the model validation set samples.

[0075] The model training step S4 includes:

[0076] Step S41: treat the unlabeled samples in the k pieces of data as negative samples, and form a new training data set with the positive samples, and use the out-of-bag data as the validation data set;

[0077] Step S42: Based on the training data set and the validation data set, k learners are used to perform training and establish k prediction models.

[0078] Step S43: For each of the k established prediction models, prediction is performed on the independent out-of-bag data in step S41, and each sample in the out-of-bag data can obtain k prediction results; wherein, the k prediction results are a probability value;

[0079] Step S44: averaging the k prediction results as the final prediction result of the out-of-bag data;

[0080] Step S45: Based on the final prediction results of the out-of-bag data, the k prediction models are subjected to model evaluation index processing and optimization processing to obtain the final k prediction models.

[0081] In an embodiment of the present invention, step S4 includes performing model evaluation index processing and tuning processing on the k prediction models (step S45); the model evaluation index includes AUC value, Log loss value and relative information gain RIG value.

[0082] Model evaluation metrics include AUC (Area Under Curve), Log Loss, and Relative Information Gain (RIG). Generally speaking, the closer the AUC is to 1, the better the model classification performance. A smaller Log Loss indicates a higher accuracy in CTR estimation; a larger RIG indicates a better model performance.

[0083] For example, after processing the data according to the above steps and training the model, the local verified AUC value can be used to judge 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:

[0084] ①. Add Batch Normalization to solve the problem of internal covariate shift of the data.

[0085] ②. Add Dropout to the network, that is, let some neurons be dormant during training.

[0086] ③. Adjust the learning rate. Generally, the learning rate during training is adjusted through strategies such as exponential decay.

[0087] ④ Set up multiple seed training and take average to reduce the risk of overfitting during training.

[0088] ⑤. Add L1 or L2 regularization to impose penalties on the loss function to reduce the risk of overfitting.

[0089] ⑥. Optimization of hyperparameters.

[0090] In terms of hyperparameter optimization methods, grid search or random search can generally be used; however, the above two methods consume a lot of computing resources and are not efficient. In an embodiment of the present invention, a Bayesian Optimization strategy is adopted. Bayesian optimization calculates the posterior probability distribution of the first n data points through Gaussian process regression to obtain the mean and variance of each hyperparameter at each value point; Bayesian optimization balances the mean and variance and ultimately selects a better set of hyperparameters based on the joint probability distribution between hyperparameters.

[0091] After all the above processing steps are completed, by introducing the features of the sample to be predicted into the trained k prediction models, it is possible to screen out users with high willingness in advance before the advertisement is delivered, and to accurately deliver marketing advertisements to these users. That is, the present invention may also include a prediction step S5, which specifically includes:

[0092] Step S51: Obtain a user group for which click prediction for a marketing campaign is to be performed and a sample of users of the user group to be predicted, and extract original feature information from the sample to be predicted; wherein the original feature information includes user ID, user province, user city, user access DPI, user access time, device type, and / or whether the user clicks; wherein the user province, user city, and user access DPI are categorical features, the user access time is a continuous feature, and N is greater than or equal to 2;

[0093] Step S52: Perform anomaly detection and processing on the raw data information of each user; process the categorical features through one-hot encoding, perform normalization and gradient descent processing according to business needs and solution model selection, or use the RankGauss method to adjust the sample distribution of continuous features;

[0094] Step S53: Provide the final k trained prediction models, input the characteristics of each user to be evaluated (sample to be predicted) into the k estimation models in turn, obtain k prediction results, and finally take the average of these k prediction results to obtain the prediction values of all the users in the user group for click prediction of the marketing activity; wherein the model prediction value of the user is the click intention of the user.

[0095] Step S54: According to actual delivery needs, select all or part of the users whose model predicted values have a concentrated click intention of 1 to perform precision marketing tasks.

[0096] The results show that through the above-mentioned feature screening scheme of the present invention, relevant experiments were carried out on certain actual click-through rate prediction data, and the original features were effectively screened by constructing shadow features; the screened features significantly reduced the overfitting risk of the tree model and improved the results of user click-through rate prediction.

[0097] The above descriptions are merely preferred embodiments of the present invention, and the embodiments are not intended to limit the scope of patent protection of the present invention. Therefore, any equivalent structural changes made using the description and drawings of the present invention should also be included in the scope of protection of the present invention.

Claims

1. A predictive ensemble modeling method based on unlabeled sample learning, characterized in that: It includes data preprocessing step S1, feature engineering processing step S2, positive / unlabeled sample random segmentation and combination ensemble training strategy step S3 and model training step S4; The data preprocessing step S1 includes the following steps: Step S11: Obtaining raw data information of N users, and extracting raw feature information from the raw data information of the users; wherein the raw feature information includes user ID, user province, user city, user access DPI, user access time (frequency), device type, and / or whether the user clicks; wherein the user province, user city, and user access DPI are categorical features, the user access time is a continuous feature, and N is greater than or equal to 2; Step S12: performing abnormality detection and processing on the original feature information of the user; Step S13: Process the categorical features through one-hot encoding, perform normalization and gradient descent processing according to business needs and solution model selection, or use the RankGauss method to adjust the sample distribution of continuous features; The feature engineering processing step S2 includes the following steps: Step S21: User access DPI processing step, creating a feature column of user ID and user access DPI; Step S22: Processing the category features by one-hot encoding; Step S23: Perform numerical feature transformation on continuous features; Step S24: performing feature screening and feature construction to form new features; Step S25: performing continuous feature discretization processing; Step S3 specifically includes the following steps: Randomly sample the positive label samples and unlabeled samples in the original training data and combine them into k new sample spaces according to a certain sample ratio; at the same time, retain a part of the data as out-of-bag data for verification; The model training step S4 includes: Step S41: treat the unlabeled samples in the k pieces of data as negative samples, and form a new training data set with the positive samples, and use the out-of-bag data as the validation data set; Step S42: Using k learners to train and establish k prediction models based on the training data set and the validation data set; Step S43: For each of the k established prediction models, prediction is performed on the independent out-of-bag data in step S41, and each sample in the out-of-bag data can obtain k prediction results; wherein, the k prediction results are a probability value; Step S44: averaging the k prediction results as the final prediction result of the out-of-bag data; Step S45: Based on the final prediction results of the out-of-bag data, the k prediction models are subjected to model evaluation index processing and optimization processing to obtain the final k prediction models.

2. The predictive integrated modeling method according to claim 1, characterized in that: The ratio of the training set samples to the validation set samples is adjusted according to the validation effect of the model validation set samples.

3. The predictive integrated modeling method according to claim 1, characterized in that: The k prediction models are gradient boosting decision tree models, and their implementation methods include XGBoost or lightGBM.

4. The predictive integrated modeling method according to claim 1, characterized in that: The model evaluation indicators include AUC value, Log loss value and relative information gain RIG value.

5. The predictive integrated modeling method according to claim 1, characterized in that: The method also includes a prediction step S5, which specifically includes: Step S51: Obtain a user group for which click prediction for a marketing campaign is to be performed and a sample of users of the user group to be predicted, and extract original feature information from the sample to be predicted; wherein the original feature information includes user ID, user province, user city, user access DPI, user access time, device type, and / or whether the user clicks; wherein the user province, user city, and user access DPI are categorical features, the user access time is a continuous feature, and N is greater than or equal to 2; Step S52: Perform anomaly detection and processing on the raw data information of each user; process the categorical features through one-hot encoding, perform normalization and gradient descent processing according to business needs and solution model selection, or use the RankGauss method to adjust the sample distribution of continuous features; Step S53: Provide the final k trained prediction models, input the characteristics of each user to be evaluated into the k estimation models in turn, obtain k prediction results, and finally take the average of these k prediction results to obtain the prediction values of all the users in the user group for click prediction of the marketing activity; wherein the model prediction value of the user is the click intention of the user.

6. The predictive integrated modeling method according to claim 5, characterized in that: The model prediction step S5 further includes: Step S54: According to actual delivery needs, select all or part of the users whose model predicted values have a concentrated click intention of 1 to perform precision marketing tasks.

Images