Semi-supervised user behavior prediction method based on DPI co-occurrence relationship features

Abstract

A semi-supervised user behavior prediction method based on DPI co-occurrence relationship features includes a data preprocessing step S1, a training set generation step S2, and a model building and prediction step S3. The present invention is a semi-supervised user behavior prediction method based on DPI co-occurrence relationship features. In order to solve the problem of feature sparsity caused by the difficulty in obtaining DPI access time and DPI one-hot encoding, it is proposed to use DPI co-occurrence conditions to construct a statistical matrix, and then perform SVD decomposition on the matrix to obtain the feature vector of each DPI; through the user's DPI features, a model is established using non-missing user attribute features to predict the missing values ​​of these features; then, the user attribute features containing the predicted values ​​and the values ​​themselves are combined with the DPI features as the user's overall features, and used to establish a click-through rate prediction model based on the user's advertising click situation to improve the user's click-through rate prediction results.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence in Internet marketing, and more specifically, to a semi-supervised user behavior prediction method based on DPI co-occurrence relationship features. 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 of 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] In click-through rate estimation, data features can generally be divided into:

[0006] ① User profile features, which mainly include the user's age, gender, income level, device type, etc.;

[0007] ② Item / commodity features, which mainly include DPI, click frequency, etc.) On this basis, features can also be interacted and combined by combining specific business needs or using algorithms, including various statistical features.

[0008] In particular, for certain product category features, after processing (e.g., one-hot vector expansion), the dimensionality is often very high, and the feature composition is relatively complex. Furthermore, in many businesses, the contribution of features to the test results is often a concern.

[0009] Therefore, during the estimation process, feature screening is beneficial in many aspects, such as extracting important features, reducing computational complexity, and reducing the risk of overfitting.

[0010] When user DPIs are relatively complete, all users access a large number of DPIs. Using one-hot encoding to expand a user's DPI accesses results in excessively high feature dimensionality, leading to significant computational overhead. Furthermore, the proportion of DPIs accessed by a single user to the total DPI accesses is relatively low, resulting in a very sparse user DPI access feature, which can lead to overfitting during modeling.

[0011] One approach to solving this problem is to collect data on the time each user accesses the DPI and sort it by the order of DPI access to obtain contextual information about the DPI. This information can then be used to train the DPI to generate feature representation vectors (i.e., word vectors). However, this approach is also very costly, requiring both DPI access information and access time, and requiring certain accuracy in access time data.

[0012] When user feature data is missing, it is usually necessary to remove the missing data before modeling to prevent the missing values ​​from affecting model building and prediction. Alternatively, missing values ​​can be artificially filled using methods such as taking the mode or mean.

[0013] The former method results in loss of original data information, as only some features are typically missing. Simply discarding user data with missing values ​​results in data waste. The latter method is overly simplistic, and when there are many missing values, it may lead to incorrect estimates of the data distribution, thus affecting modeling and prediction. Summary of the Invention

[0014] The purpose of the present invention is to provide a semi-supervised user behavior prediction method based on DPI co-occurrence relationship features. It constructs a statistical matrix using DPI co-occurrence conditions, and then performs SVD decomposition on the matrix to obtain the feature vector of each DPI as a feature word vector to improve the user click-through rate prediction result.

[0015] To achieve the above object, the technical solution of the present invention is as follows:

[0016] A semi-supervised user behavior prediction method based on DPI co-occurrence relationship features, characterized by comprising a data preprocessing step S1, a training set generation step S2, and a model training and model establishment step S3;

[0017] The data preprocessing step S1 includes the following steps:

[0018] Step S11: obtaining raw data information of m 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, m is greater than or equal to 2; the number of user access DPIs is n, and n is greater than or equal to 2;

[0019] Step S12: performing abnormality detection and processing on the original feature information of the user;

[0020] 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;

[0021] The training set generation step S2 includes a feature engineering processing step S21, a co-occurrence matrix construction and DPI feature vector acquisition step S22, and a model training and model establishment step S23;

[0022] Step S21 specifically includes the following steps:

[0023] Step S211: User access DPI processing step, creating a feature column of user ID and user access DPI;

[0024] Step S212: Processing the category features by one-hot encoding;

[0025] Step S213: Perform numerical feature transformation on continuous features;

[0026] Step S214: performing feature screening and feature construction to form new features;

[0027] Step S215: performing continuous feature discretization processing;

[0028] Step S22: The construction of the co-occurrence matrix and the acquisition of the DPI feature vector specifically include the following steps:

[0029] Step S221: Obtain information about all users accessing the DPI on the same day through the user access DPI access list, and then count the users in the user access DPI access list to obtain a co-occurrence matrix; wherein the dimension of the co-occurrence matrix is ​​n×n;

[0030] Step S221: The co-occurrence matrix can be decomposed using the singular value decomposition algorithm to obtain a feature matrix with dimensions of n×k, where k is the dimension of the feature matrix (k < n). The dimension of the feature matrix is freely specified and represents the feature dimension. For each DPI, there is a corresponding k-dimensional feature vector, and the k-dimensional feature vector is the dense representation of the user accessing the DPI, that is, the word vector.

[0031] Step S3 specifically includes the following steps:

[0032] Step S311: Obtain the feature vector of each DPI through the co-occurrence matrix. Then, for each user, take the mean of the feature vectors corresponding to the DPI accessed by the user to obtain the k-dimensional DPI feature.

[0033] Step S312: Screen out the users with non-missing user attribute features. Use the DPI feature and the user attribute feature to establish a user attribute prediction model, and then use the DPI feature of the users missing the user attribute feature to bring into the model to predict the missing user attribute features of these users.

[0034] Step S313: Combine the DPI features and user attribute features of all users, and combine the advertising click label of each user to establish a final click-through rate prediction model.

[0035] Step S32: By bringing the above features into the trained user prediction model, it is possible to screen out some users with higher willingness in advance before advertising is launched and perform precise delivery of marketing advertisements to these users.

[0036] Furthermore, the click-through rate prediction model is a gradient boosting decision tree model, and its implementation methods include XGBoost or lightGBM.

[0037] From the above technical solutions, it can be seen that the semi-supervised user behavior prediction method based on the DPI co-occurrence relationship features of the present invention aims at the problems of difficult access to DPI access time and feature sparsity caused by DPI one-hot encoding. It proposes to use the DPI co-occurrence situation to construct a statistical matrix, and then perform SVD decomposition on this matrix to obtain the feature vector of each DPI as the feature word vector.

[0038] For each user, the DPI feature of the user is the mean vector of the word vectors of the user accessing each DPI. This DPI feature has a specific dimension, which can be freely specified and selected by algorithm developers according to the modeling and testing situations.

[0039] Since the DPI access behavior of users reflects the characteristics and attributes of users to a certain extent, there is a certain connection between some user attribute features such as age, gender, and consumption behavior of users and the DPI access situation.

[0040] However, user attribute features are often difficult to obtain, and these features may contain a large number of missing values. This paper uses the user's DPI features to build a model using non-missing user attribute features to predict missing values ​​for these features. The user attribute features containing both predicted values ​​and actual values, along with the DPI features, are then combined as overall user features to build a click-through rate prediction model based on the user's ad clicks.

[0041] That is to say, since the present invention uses the user attribute features predicted by the user DPI features to fill in the missing values, and these values ​​are also used to establish the click-through rate prediction model, the established click-through rate prediction model is a semi-supervised model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The figure shows a flow chart of a semi-supervised user behavior prediction method based on DPI co-occurrence relationship features in an embodiment of the present invention.

[0043] Figure 2 FIG2 is a schematic diagram showing the process of constructing a co-occurrence matrix using the user DPI access situation and obtaining a DPI feature vector through the constructed collinear matrix according to an embodiment of the present invention.

[0044] Figure 3 The figure shows the establishment of the missing feature prediction and click rate estimation model in the embodiment of the present invention.

[0045] Figure 4 The figure shows the DPI features constructed by the co-occurrence matrix in an embodiment of the present invention, and the binary classification prediction of whether the user clicked 10 times, and the schematic diagram of the probability density distribution on the training set and test set. DETAILED DESCRIPTION

[0046] The following is combined with Figure 1-4 , the specific implementation methods of the present invention are further described in detail.

[0047] 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.

[0048] See also Figure 1 , Figure 1 The figure shows a flow chart of the semi-supervised user behavior prediction method based on DPI co-occurrence relationship features of the present invention. Figure 1 As shown, the prediction method includes a data preprocessing step S1, a training set generation step S2, and a model building and prediction step S3.

[0049] 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:

[0050] Step S11: Take the original data information of m 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 province (province), user city (city), user access DPI, user access time (frequency), device type (phone_type) and / or whether the user clicks on the feature; wherein the user province, user city and user access DPI are category features, the user access time is a continuous feature, and m is greater than or equal to 2; the number of user access DPIs is n, and n is greater than or equal to 2.

[0051] 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:

[0052] Table 1

[0053]

[0054] 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.

[0055] 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.

[0056] Step S12: performing abnormality detection and processing on the original data information of the user.

[0057] 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.

[0058] 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.

[0059] 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.

[0060] 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.

[0061] 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.

[0062] Step S2: The step of generating a training set, which includes a feature engineering processing step S21, a shadow feature construction step S22, and a feature screening step S23 based on feature shadows.

[0063] Step S21: 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:

[0064] Step S211: 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.

[0065] Step S212: One-hot encoding is used to process categorical features. Categorical features such as the province, city, and device type of the user number 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 user is 1, and the provinces of other users are 0. For example, user ID 199990 has the province HN, so after one-hot encoding, its location vector is (1, 0, 0, 0, 0); correspondingly, the location vector of user 199991 is (0, 1, 0, 0, 0).

[0066] Step S213: Perform numerical feature transformation processing on continuous features.

[0067] 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.

[0068] Step S214: Perform feature screening and feature construction to form new features.

[0069] 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.

[0070] Feature construction: Add commonly used statistical features such as skewness, peakedness, etc.

[0071] Step S215: Discretize continuous features.

[0072] 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.

[0073] After preprocessing and feature engineering, the data is available as shown in Table 2 below:

[0074]

[0075]

[0076] Continuous features include statistical features such as DPI access frequency, kurtosis, and skewness. Categorical features include the one-hot encoding of DPI access, the one-hot encoding of the number's location, and the discretized DPI access frequency.

[0077] Step S22: constructing a co-occurrence matrix and obtaining a DPI feature vector;

[0078] Please refer to Figure 2 , Figure 2 The figure shows a schematic diagram of the process of constructing a co-occurrence matrix using the user DPI access situation and obtaining the DPI feature vector through the constructed collinear matrix in the embodiment of the present invention. As Figure 2 shown, the situation of the user accessing the DPI on the same day can be obtained through the DPI access list, and then the DPI in the list is counted to obtain the co-occurrence matrix. The figure shows the co-occurrence situation of three users accessing the DPI: the co-occurrence of user accessing DPI A, user accessing DPI B, and user accessing DPI C. Among them, 14 means that user accessing DPI A and user accessing DPI B appear together 14 times in the training data of all users, that is, there are 14 users in the data who access DPI A and DPI B at the same time. The other data in the matrix can be inferred by analogy.

[0079] Suppose there are n users accessing the DPI in total, then the dimension of the co-occurrence matrix is n×n. After obtaining the co-occurrence matrix, the singular value decomposition algorithm can be used to decompose the co-occurrence matrix to obtain a feature matrix with the dimension of n×k, where k is the dimension of the feature matrix (k < n), and this dimension can be freely specified, representing the feature dimension. For each DPI, there is a corresponding k-dimensional feature vector, and this feature vector is the dense representation of the user accessing the DPI, that is, the word vector. After the above data processing steps, the newly generated data is divided into a training set and a validation set, and the model building step S3 is carried out.

[0080] Step S31: Establish a neural network initialization model based on the tree algorithm.

[0081] Please refer to Figure 3 , Figure 3 The figure shows a schematic diagram of the establishment of the missing feature prediction and click-through rate prediction model in the embodiment of the present invention. As Figure 3 shown, in the process of predicting the missing features and establishing the final click-through rate prediction model, the following steps are carried out:

[0082] Step S311: Obtain the feature vector of each DPI through the co-occurrence matrix. Then, for each user, the feature vectors corresponding to the DPI accessed by the user are averaged to obtain the k-dimensional DPI feature.

[0083] Step S312: Screen out the users with non-missing user attribute features, establish a user attribute prediction model using the DPI feature and the user attribute feature, and then use the DPI feature of the users lacking the user attribute feature to bring into the model to predict the missing user attribute features of these users.

[0084] Step S313: Merge the DPI features and user attribute features of all users, combine them with each user's ad click label, and establish a final click rate prediction model.

[0085] In an embodiment of the present invention, after having the above-mentioned click-through rate prediction model, it is possible to bring the above-mentioned features into the trained user prediction model to screen out some users with higher willingness in advance before the advertisement is delivered, and accurately deliver marketing advertisements to these users, that is, prediction step S32.

[0086] See also Figure 4 , Figure 4 The figure shows the DPI features constructed by the co-occurrence matrix in the embodiment of the present invention, and the binary classification prediction of whether the user clicked 10 times, and the schematic diagram of the probability density distribution on the training set and the test set. Figure 4 As shown in the figure, the DPI features constructed using the co-occurrence matrix can be used to perform a binary classification prediction of whether a user clicked on 10 messages. The probability density distributions on the training and test sets are almost identical, effectively suppressing overfitting. Furthermore, the AUC value on the test set is 0.83, indicating that the model has good ability to distinguish between positive and negative samples.

[0087] The results show that the feature screening scheme of the present invention effectively realizes the feature construction based on the DPI co-occurrence matrix, and significantly improves the results of user click-through rate prediction.

[0088] 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 semi-supervised user behavior prediction method based on DPI co-occurrence relationship features, characterized by: It includes a data preprocessing step S1, a training set generation step S2, and a model establishment and prediction step S3; The data preprocessing step S1 includes the following steps: Step S11: Obtain the original data information of m users, and extract the original feature information from the original data information of the users; wherein, the original feature information includes user ID, user's province, user's city, user's access DPI, user's access time, device type, and / or whether the user clicks the feature; wherein, the user's province, user's city, and user's access DPI are categorical features, the user's access time is a continuous feature, and m is greater than or equal to 2; the number of user access DPIs is n, and n is greater than or equal to 2; Step S12: Perform an anomaly detection and processing step on the original feature information of the users; Step S13: Process the categorical features through one-hot encoding, and perform normalization and then gradient descent processing according to business requirements and the selected solution model, or use the RankGauss method to adjust the sample distribution of the continuous features; The step S2 includes a feature engineering processing step S21, a co-occurrence matrix construction and DPI feature vector acquisition step S22, and a model training and model establishment step S23; Step S21 specifically includes the following steps: Step S211: A user access DPI processing step to establish a feature column of user ID and user access DPI; Step S212: Use one-hot encoding to process the categorical features; Step S213: Perform a numerical feature transformation process on the continuous features; Step S214: Perform feature screening and feature construction to form new features; Step S215: Perform continuous feature discretization processing; The construction of the co-occurrence matrix and the acquisition of the DPI feature vector in step S22 specifically include the following steps: Step S221: Through the user access DPI access list, obtain the situation of all users accessing the user access DPI on the same day, and then count the user access DPIs in the user access DPI access list to obtain the co-occurrence matrix; wherein, the dimension of the co-occurrence matrix is n×n; Step S221: The co-occurrence matrix can be decomposed using the singular value decomposition algorithm to obtain a feature matrix with a dimension of n×k, where k is the dimension of the feature matrix (k < n), the dimension of the feature matrix is freely specified, representing the feature dimension, and for each DPI, there is a corresponding k-dimensional feature vector, and the k-dimensional feature vector is the dense representation of the user access DPI, that is, the word vector; Step S3 specifically includes the following steps: Step S31: Obtain the feature vector of each DPI through the co-occurrence matrix. Then, for each user, take the mean of the feature vectors corresponding to the DPIs it accesses to obtain the k-dimensional DPI feature; Step S32: Screen out the users with non-missing user attribute features, use the DPI features and user attribute features to establish a user attribute prediction model, and then bring the DPI features of the users missing user attribute features into the model to predict the missing user attribute features of these users; Step S33: Merge the DPI features and user attribute features of all users, combine them with each user's ad click label, and establish a final click rate prediction model.

2. The prediction method according to claim 1, characterized in that The click-through rate prediction model is a gradient boosting decision tree model, and its implementation methods include XGBoost or lightGBM.

3. The prediction method according to claim 1, wherein: The step S3 also includes a step S32: by introducing the above features into the trained user prediction model, it is possible to screen out some users with higher willingness in advance before the advertisement is delivered, and accurately deliver marketing advertisements to these users.

Images