Advertisement marketing modeling method based on financial field and related equipment

Abstract

The invention discloses an advertisement marketing modeling method and device based on the financial field, a server and a computer readable storage medium. According to the method, samples are divided into a plurality of subsets according to classification performance of models on sample prediction results, corresponding feature extraction models are trained in a targeted mode in advance for different types of subsample sets, the feature extraction models are in one-to-one correspondence with the subsample sets, each feature extraction model independently learns features of the corresponding subsample set, and the feature extraction efficiency is improved. Each feature extraction model can carry out better feature learning on sub-sample sets with balanced distribution, interference of uneven overall distribution of the sample sets on model recognition can be avoided, a better recognition and classification effect can be achieved, feature values generated by the feature extraction models are fused, and the recognition and classification efficiency is improved. And calling the pre-trained evaluation model to carry out user demand prediction on the financial product based on the fused feature value, and realizing accurate user demand prediction based on the accurate fused feature value.

Description

Technical Field

[0001] The present disclosure generally relates to the financial field, and specifically relates to an advertising and marketing modeling method, device, server, and computer-readable storage medium based on the financial field. Background Art

[0002] Advertising marketing is a strategy to deliver product or service information to target audiences through various media channels, with the goal of increasing brand awareness, promoting sales, and establishing connections with potential customers. Advertising marketing algorithms refer to a method of calculating the user's willingness to demand a product by learning and encoding the information of potential users. In the financial field, advertising marketing models are often used to determine a potential user's demand for financial products.

[0003] The current advertising and marketing algorithm mainly models the user's input features through the LR model, assigns weights to the user's features, and outputs the weighted sum of the features to obtain a probability score, indicating the user's demand for the financial product.

[0004] However, in the field of financial advertising, user data is usually sparse, feature coverage is low, and the overall distribution of samples is uneven, resulting in poor modeling effect, low prediction accuracy, and poor credibility. Summary of the invention

[0005] In view of the above-mentioned defects or deficiencies in the prior art, it is desired to provide an advertising and marketing modeling method, device, service and computer-readable storage medium that can significantly improve the accuracy of user demand assessment for financial products.

[0006] In a first aspect, the present application embodiment provides an advertising marketing modeling method based on the financial field, including:

[0007] Obtain the collected user feature data as a sample set;

[0008] Dividing the sample set into sample types according to the classification performance of the samples to obtain a plurality of sub-sample sets;

[0009] Calling the feature extraction model corresponding to each sample type, extracting features from the corresponding sub-sample set, and fusing the groups of feature values ​​generated by each of the feature extraction models;

[0010] The evaluation model is called to predict the demand for the financial product based on the fused feature values, and the prediction result is used as the user's demand for the financial product.

[0011] In one embodiment, dividing the sample set into sample types according to the classification performance of the samples includes:

[0012] Calling a pre-trained feature extraction model to perform sample evaluation on the sample set;

[0013] Determine the dividing line between positive and negative samples in the evaluation results;

[0014] Extracting a first proportion of samples from the sample set as a difficult sample set according to the distance from the segmentation line from near to far;

[0015] According to the distance from the segmentation line from far to near, samples of a second proportion are extracted from the sample set as a noise sample set.

[0016] In one embodiment, calling the feature extraction model corresponding to each sample type, performing feature extraction on the corresponding sub-sample set, and fusing each set of feature values ​​generated by each feature extraction model includes:

[0017] Calling the first feature extraction model to extract features from each sample in the difficult sample set, and using the generated features as the first features retrieved;

[0018] Eliminating the noise sample set from the sample set, and calling the second feature extraction model to extract features from the remaining samples, and using the generated features as the second features after eliminating interference;

[0019] Calling a third feature extraction model to extract features from each sample in the sample set, and using the generated features as the basic third features;

[0020] Feature fusion is performed on the first feature, the second feature, and the third feature.

[0021] In one embodiment, before dividing the sample set into sample types according to the classification performance of the samples, the method further includes:

[0022] Determine the missing core features in the user feature data as target features;

[0023] Calling a pre-trained feature prediction model to predict missing data for the target feature, and backfilling the predicted data into the user feature data;

[0024] The backfilled user feature data is used as the sample set.

[0025] In one embodiment, determining that there are missing core features in the user feature data includes:

[0026] Input the user feature data into the LR model for evaluation to obtain the weight corresponding to each feature;

[0027] Extract the first M user features in order according to the weights;

[0028] Determine the coverage of each feature in the first M user features;

[0029] The features with coverage rates higher than a threshold are removed from the first M user features, and the remaining features are used as the target features.

[0030] In one embodiment, before dividing the sample set into sample types according to the classification performance of the samples, the method further includes:

[0031] Determining key features in the user feature data;

[0032] Performing feature cross processing on the key features to generate new features;

[0033] The user feature data and the new feature are combined as a sample set.

[0034] In one embodiment, determining the key features in the user features includes:

[0035] Input the user feature data into the LR model for evaluation to obtain the weight corresponding to each feature;

[0036] According to the weights, the first N user features are extracted in order;

[0037] Determine whether there is a gap in the weights of the first N user features;

[0038] If there is a fault, all user features before the fault are used as the key features;

[0039] If there is no fault, the first N user features are used as the key features.

[0040] In a second aspect, the embodiment of the present application provides an advertising and marketing modeling device based on the financial field, including:

[0041] A data acquisition unit, used to acquire the collected user feature data as a sample set;

[0042] A sample set division unit, used to divide the sample set into sample types according to the classification performance of the samples to obtain a plurality of sub-sample sets;

[0043] A multi-model extraction unit, used to call the feature extraction model corresponding to each sample type, perform feature extraction on the corresponding sub-sample set, and fuse each set of feature values ​​generated by each of the feature extraction models;

[0044] The demand assessment unit is used to call the assessment model to predict the demand for the financial product based on the fused feature value, and use the prediction result as the user's demand for the financial product.

[0045] In a third aspect, an embodiment of the present application provides a server, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the embodiment of the present application when executing the program.

[0046] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the embodiment of the present application.

[0047] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0049] Figure 1 A schematic diagram of a process of an advertising and marketing modeling method based on the financial field provided in an embodiment of the present application is shown;

[0050] Figure 2 A schematic diagram of task flow provided by an embodiment of the present application is shown;

[0051] Figure 3 An exemplary structural block diagram of an advertising and marketing modeling device based on the financial field provided by an embodiment of the present application is shown;

[0052] Figure 4 A schematic diagram of the structure of a computer system suitable for implementing a server of an embodiment of the present application is shown. DETAILED DESCRIPTION

[0053] The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0054] It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments. Although the embodiments of the present application provide the method operation instruction steps shown in the following embodiments or drawings, more or fewer operation instruction steps may be included in the method based on routine or no creative labor. In the steps where there is no necessary causal relationship logically, the execution order of these steps is not limited to the execution order provided by the embodiments of the present application. The method may be executed in the order of the methods shown in the embodiments or drawings or in parallel during the actual processing process or when the device is executed.

[0055] It should be noted that the acquisition or use of data in the embodiments of the present application requires the user's consent. The relevant data can only be obtained after the user's authorization and permission, and the acquisition or use of the data complies with relevant legal provisions.

[0056] Please refer to Figure 1 , Figure 1 The flowchart of the advertising and marketing modeling method based on the financial field provided by an embodiment of the present application is shown. Figure 1 As shown, the method includes:

[0057] Step S101, obtaining collected user feature data as a sample set;

[0058] Obtain user characteristic data for forecasting demand for financial products, such as age, income, education, social information, credit information, etc. It should be noted that the acquisition or use of the above information requires the user's consent, and the above data can only be obtained after the user's authorization.

[0059] Step S102: dividing the sample set into sample types according to the classification performance of the samples to obtain a number of sub-sample sets;

[0060] Step S103: calling the feature extraction model corresponding to each sample type, performing feature extraction on the corresponding sub-sample set, and fusing each set of feature values ​​generated by each feature extraction model;

[0061] Here, step S102 and step S103 are combined for introduction. At present, in modeling, all data are input into the model for training and learning. However, since the data of users in the field of financial advertising is usually sparse, the overall distribution of samples is uneven, and the model focuses on the overall distribution of samples, the modeling effect of the model is poor and the prediction result is low in accuracy. In order to solve the problem of low model recognition accuracy caused by user data, this method proposes to divide the sample set into sample types according to the classification performance of the samples, and then call the feature extraction models trained specifically according to each sample type to extract features from the corresponding sub-sample sets.

[0062] Among them, the classification performance of the sample refers to whether the model predicts that the sample is a positive sample (a user who has used the specified financial product) or a negative sample (a user who has not used the specified financial product) after the sample is input into the model, or whether it is difficult to distinguish between positive and negative samples. A normal model can output a higher proportion of predicted positive data, such as 99%, for data that is actually a positive sample, but if the model effect is not good, it may output a lower proportion of predicted positive data, such as 52%, or even predict negative data that is actually a positive sample. In this method, the sample is divided into different subsets according to the classification performance of the model for the sample, such as a sample set with a model recognition accuracy of 80% to 100%, a sample set with a model recognition accuracy of 50% to 80%, a sample set with model recognition errors, a sample set with model recognition difficulties, etc. The specific division of the sample set can be set according to different recognition accuracy requirements and available resources in actual applications. In this embodiment, only the above subsets are used as examples, and other divisions can refer to the introduction of this embodiment. It should be noted that in this method, samples between sub-sample sets are allowed to be cross-collected, for example, sample 1 can be in sub-sample set 1 and sub-sample set 2 at the same time.

[0063] For different types of sub-sample sets, the corresponding feature extraction models are pre-trained in a targeted manner. For example, feature extraction model 1 is trained first, and for sample sets with recognition accuracy of model 1 between 80% and 100%, model 1 is still selected as the corresponding feature extraction model; for sample sets with recognition accuracy of model 1 between 50% and 80%, model 2 is trained for this part of the sample set as the corresponding feature extraction model; model 3 is trained for sample sets with recognition errors of model 1; model 4 is trained for sample sets with difficulty in recognition of model 1, etc. The feature extraction model corresponds to the sub-sample set one by one, and each feature extraction model learns the features of the corresponding sub-sample set separately, which can avoid the interference of the overall uneven distribution of the sample set on the model recognition. For each feature extraction model in this method, the corresponding sub-sample set is evenly distributed and has obvious features. Therefore, each feature extraction model can perform better feature learning for the evenly distributed sub-sample set, so as to have a better recognition and classification effect.

[0064] In the above steps, there is no limit on the number of sub-sample sets divided, and the feature extraction model is set corresponding to the sub-sample set, so there is no limit on the number of feature extraction models. In addition, the specific model type and model structure parameters of the feature extraction model are not limited in this embodiment, and the model selection can be referenced according to relevant technologies.

[0065] After calling the feature extraction model corresponding to each sample type and extracting features from the corresponding sub-sample set, each feature extraction model will output a set of feature values. The feature values ​​generated by each feature extraction model are fused, and the fused feature values ​​contain the features of all samples in the overall sample set. The feature value fusion method can refer to the implementation of related technologies and is not limited here.

[0066] Step S104: Call the evaluation model to predict the demand for the financial product based on the fused feature value, and use the prediction result as the user's demand for the financial product.

[0067] After obtaining the fused feature values, the pre-trained evaluation model is called to predict user demand for financial products based on the fused feature values. The specific model type and model structure parameters of the evaluation model are not limited. For example, a linear regression model, a logistic regression model, a decision tree, etc. can be selected, and the prediction result (i.e., probability value) output by the model is output as the user's demand for the financial product. The higher the model score, the higher the possibility that the user needs the financial product.

[0068] It should be noted that although the operations of the method of the present invention are described in a particular order in the drawings, this does not require or imply that the operations must be performed in this particular order or that all illustrated operations must be performed to achieve desired results.

[0069] Based on the above introduction, the method provided in this embodiment divides the samples into several subsets according to the classification performance of the model for the sample prediction results. For different types of sub-sample sets, the corresponding feature extraction models are pre-trained in a targeted manner. The feature extraction models correspond to the sub-sample sets one by one. Each feature extraction model learns the features of the corresponding sub-sample set separately. For each feature extraction model, the corresponding sub-sample sets are evenly distributed and have obvious features. Therefore, each feature extraction model can perform better feature learning for the evenly distributed sub-sample sets, which improves the expression ability of each model and avoids the interference of the overall uneven distribution of the sample sets on the model recognition, so as to have a better recognition and classification effect. The feature values ​​generated by each feature extraction model are then fused, and the pre-trained evaluation model is called to perform user demand prediction for financial products based on the fused feature values. Accurate user demand prediction can be achieved based on accurate fused feature values.

[0070] In this embodiment, a specific method for dividing the sample set into sample types according to the classification performance of the samples is proposed for step S102, which specifically includes the following sub-steps:

[0071] Step S21, calling the pre-trained feature extraction model to perform sample evaluation on the sample set;

[0072] Train a feature extraction model (such as the XGBOOST model (extreme gradient boosted tree model)) for all the data. This model can be understood as a large and comprehensive model that can learn the distribution of samples from a global perspective. After training the model, call the model to score the samples in the sample set.

[0073] Step S22, determining the dividing line between positive and negative samples in the evaluation results;

[0074] Determine the dividing line between positive and negative samples in the evaluation results. Optionally, you can find the dividing points corresponding to KS (Kolmogorov-Smirnov, used to evaluate the discrimination of the model) and determine the dividing line composed of the dividing points corresponding to KS. This dividing line is the best dividing line for positive and negative samples that the model considers. The dividing line can also be determined by other indicators, which will not be repeated here.

[0075] Ideally, all positive samples are distributed above the dividing line, and all negative samples are distributed below the dividing line. Misclassified positive samples will appear below the dividing line, and misclassified negative samples will appear above the dividing line.

[0076] Step S23, extracting a first proportion of samples from the sample set according to the distance segmentation line from near to far as a difficult sample set;

[0077] According to the distance from the dividing line from near to far, the first proportion of misclassified samples closest to the dividing line are regarded as difficult samples that the basic model does not learn well, and the first proportion is, for example, 10%.

[0078] Step S24: extract samples of a second proportion from the sample set according to the distance segmentation line from far to near as a noise sample set.

[0079] In the real world, users' real needs are constantly changing. At the same time, due to some uncontrollable factors of individuals, there will be deviations between real needs and actual label feedback, which will also affect the machine learning model. To avoid this problem, in this embodiment, noise samples are extracted. Specifically, the selection of noise samples is exactly the opposite of difficult samples. According to the distance from the dividing line from far to near, the second proportion of misclassified samples farthest from the dividing line are regarded as noise samples that the basic model cannot learn well. The second proportion is, for example, 5%.

[0080] Accordingly, in view of the above-mentioned method of dividing the sample set, this embodiment further proposes a method of extracting features from each sub-sample set. Specifically, step S103 calls the feature extraction model corresponding to each sample type, extracts features from the corresponding sub-sample set, and fuses each group of feature values ​​generated by each feature extraction model, which may include the following sub-steps:

[0081] Step S31, calling the first feature extraction model to extract features from each sample in the difficult sample set, and using the generated features as the first features retrieved;

[0082] A first feature extraction model is specially trained for difficult samples to retrieve positive samples near the boundary values. The first feature extraction model may be an XGBOOST model, or other types of models may be used, which are not limited here.

[0083] Step S32: remove the noise sample set from the sample set, and call the second feature extraction model to extract features from the remaining samples, and use the generated features as the second features after eliminating interference;

[0084] Find the noise sample, remove the sample and then train a second feature model. The second feature model can be understood as a supplementary model. Calling the second feature model to extract features from the sample set that has removed the noise sample set can eliminate potential interference factors.

[0085] The second feature extraction model may also use the XGBOOST model, which is not limited here.

[0086] Step S33: calling the third feature extraction model to extract features from each sample in the sample set, and using the generated features as the basic third features;

[0087] The third feature extraction model is then trained as a basic model to extract features from all samples in the sample set, to analyze the sample set as a whole, and to extract overall distribution features.

[0088] To simplify the steps, the third feature extraction model may directly select the feature extraction model used for sample evaluation in step S21.

[0089] Step S34: perform feature fusion on the first feature, the second feature and the third feature.

[0090] The obtained eigenvalues ​​are also divided into three parts, namely, the first feature generated by the first feature extraction model as the salvage model, the second feature generated by the second feature extraction model as the noise model, and the third feature generated by the third feature extraction model as the basic model. There is no coupling between the three models. By fusing the three sets of eigenvalues, the fused overall features can be obtained.

[0091] The method provided in this embodiment selectively selects a batch of samples that are difficult for the model to learn based on the cut-off point of KS, finds difficult samples on the decision boundary through the scoring of the basic model, and constructs a salvage model. The salvage ability of the model can be improved by learning the difficult samples. In order to eliminate the interference of other factors, noise samples far away from the decision boundary are also removed, possible noise samples are eliminated, and an anti-interference model is trained to further improve the expression ability of the model.

[0092] In the field of financial advertising, user data is usually sparse, and most user information is incomplete. When the model is used for demand forecasting, the modeling effect is poor. To avoid this problem, this embodiment proposes a missing backfilling method to improve the comprehensiveness of user information.

[0093] In one embodiment, before executing step S102 to classify the sample set into sample types according to the classification performance of the samples, the following steps may be performed first:

[0094] Step S105: determining the missing core features in the user feature data as target features;

[0095] In the field of financial advertising, user data is usually sparse and feature coverage is low. In order to solve the problem of missing data, this embodiment proposes to first fill in the missing data before feature extraction. Specifically, first find the core features with high importance, because the cost of pre-training is very high, and it is not necessary to pre-train for every missing feature.

[0096] Optionally, a specific implementation step of determining the missing core features in the user feature data is as follows:

[0097] Step S51: input the user feature data into the LR (Logistic Regression) model for evaluation to obtain the weight corresponding to each feature;

[0098] All features are input into the LR model for training, and the user's real needs are used as labels for the model to learn. Because LR is a linear model, the weight of the model reflects the importance of each feature, so the weight of each feature can be obtained.

[0099] Step S52, extracting the first M user features in order according to the weights;

[0100] Wherein, M is a preset positive integer, and the specific numerical setting is not limited in this embodiment. For example, the first ten user features can be selected.

[0101] Step S53: determining the coverage rate of each feature in the first M user features;

[0102] Step S54: remove features with coverage rates higher than a threshold from the first M user features, and use the remaining features as target features.

[0103] Features with feature coverage greater than a threshold (such as 70%) are removed, because the pre-training gain of these high-coverage features is not expected to be too large.

[0104] Of course, if there are sufficient resources available, in order to further improve the accuracy, the value of M can be set relatively large, and steps S53 and S54 can be omitted, and the first M items can be directly extracted as target features, or even all user feature data can be directly used as target features, which is not limited here.

[0105] Step S106: Call the pre-trained feature prediction model to predict the missing data of the target feature, and backfill the predicted data into the user feature data;

[0106] Train a feature model to score the data with missing features, use other known features to predict the missing feature values, and fill in the scores for the missing features. For example, if the user's income data is missing, you can use features such as age and education level to predict income.

[0107] The specific model type that can be selected for the feature prediction model is not limited in this embodiment. Optionally, for continuous features, a regression model can be selected; for discrete features, a classification model can be selected. In addition, the pre-training process of the feature prediction model is not limited in this embodiment. The target feature can be used as output, and all other features can be used as input for pre-training.

[0108] Step S107: Use the backfilled user feature data as a sample set.

[0109] The method provided in this embodiment pre-trains important features, predicts missing features, obtains predicted values ​​of each missing important feature, and uses the predicted values ​​as real features for training, which can significantly improve feature coverage.

[0110] The features in the collected user feature data may have tens of thousands of dimensions, but there may be only dozens of them that can better reflect user needs. In order to further improve the expression ability of the evaluation model, this embodiment proposes a feature augmentation method to improve the richness of the sample by combining key features.

[0111] In one embodiment, before executing step S102 of dividing the sample set into sample types according to the classification performance of the samples, the following steps may be performed first:

[0112] Step S108: determining key features in the user feature data;

[0113] It should be noted that the definition criteria of the key features in this step and the core features in step S105 are not limited, and the same selection method or different methods can be selected. Since the two features have different uses, the core features in step S105 are used to backfill missing values, and the key features in this step are used to augment features. In order to achieve different purposes, two different selection methods are provided in this embodiment and the above embodiments. The specific implementation of other selection criteria can refer to this embodiment and will not be repeated here.

[0114] Optionally, a specific implementation method for determining key features in user features is as follows:

[0115] Step S81: Input the user feature data into the LR model for evaluation to obtain the weight corresponding to each feature;

[0116] This step may refer to step S51 and will not be described in detail here.

[0117] Step S82, extracting the first N user features in order according to the weights;

[0118] Wherein, N is a preset positive integer, and its specific value is not limited in this embodiment. For example, N can be 20.

[0119] Step S83, determine whether there is a gap in the weights of the first N user features; if there is a gap, use all user features before the gap as key features; if there is no gap, use the first N user features as key features.

[0120] Taking N as 20 as an example, the features are sorted according to their importance. If there is an importance gap among the top 20 features, the features before the gap are taken as important features. Otherwise, the top 20 features are taken as important features.

[0121] Among them, the method for judging the fault is not limited in this embodiment, and it can be judged by indicators such as gradient, weight, loss function, etc., which will not be repeated here.

[0122] Step S109, performing feature cross processing on key features to generate new features;

[0123] For the key features obtained, we can continue to multiply the features in pairs to get the product of the features, and then add the new features obtained to the original features to enhance the richness of the features.

[0124] Step S110: Merge the user feature data and the new features as a sample set.

[0125] The method provided in this embodiment combines key features to adaptively obtain augmented features as supplementary features, which can significantly enhance the richness of features.

[0126] It should be further explained that the above embodiments can be used in combination. To deepen understanding, this embodiment proposes a method to enhance the performance of the evaluation model from both the feature and sample perspectives. Figure 2 The following is a task flow diagram. A specific implementation step is as follows:

[0127] Step 1: Find important features in user feature data

[0128] First, all features are input into the LR model for training, and the user's real needs are used as labels for the model to learn. Because LR is a linear model, the weight of the model reflects the importance of each feature, so the weight of each feature can be obtained.

[0129] The features are sorted according to their importance. If there is an importance gap among the top 20 features, the features before the gap are taken as important features. Otherwise, the top 20 features are taken as important features.

[0130] Step 2: Pre-training of data

[0131] The problem to be solved in this step is the missing features of the data. Before supplementing the features, we must first find the features with high importance, because the cost of pre-training is very high, and there is no need to pre-train for every missing feature.

[0132] The top 10 features in terms of feature importance are selected as the pre-selected features that need to be pre-trained, and then the features with feature coverage greater than 70% are eliminated.

[0133] Finally, the remaining features are used as output, and all other features are used as input for pre-training. For continuous features, you can choose to use a regression model for training; for discrete features, you can choose a classification model for training.

[0134] The final model then scores the data with missing features and fills the missing features with the scores.

[0135] Step 3: Feature enhancement of samples

[0136] The goal of this step is to add the results obtained in the above two steps to the features of the sample to enhance the richness of the features.

[0137] For the important features obtained in the first step, we can continue to multiply the features in pairs to get the product of the features, and then add the obtained features to the original features.

[0138] So the features we get now are divided into three parts: one is the original features, another is the combined features obtained by multiplying the important features, and another is the supplementary features obtained through pre-training. The features mentioned below refer to the union of these three features.

[0139] Step 4: Selection of difficult samples

[0140] The goal of this step is to find difficult samples and then train a recovery model specifically for difficult samples.

[0141] In order to obtain difficult samples, we first need to train an XGBOOST model on all the data. After training the model, we score the samples in the training set and find the split point corresponding to KS. The 10% of misclassified samples closest to the split line are regarded as difficult samples that the basic model does not learn well.

[0142] Finally, these difficult samples are used to train an XGBOOST model, which is specifically used to retrieve the positive samples near the boundary values.

[0143] Step 5: Selection of noise samples

[0144] The goal of this step is to find noise samples, remove them, and then train an XGBOOST model, which can be understood as a complementary model.

[0145] The 5% misclassified samples farthest from the segmentation line are regarded as noise samples that the basic model cannot learn well.

[0146] After removing the noise samples, an XGBOOST model is trained. The scoring of this model excludes potential interference factors.

[0147] Step 6: Determine the edges and their characteristics

[0148] After obtaining the above three models, namely a basic model, a salvage model, and a noise model, because there is no coupling between these three models, a LR model is finally trained as the final main model. The score of the main model reflects the user's demand for the financial product.

[0149] In this solution, we enhance the performance of the XGBOOST model from two perspectives: feature and sample selection. First, by pre-training important features, we get the predicted value of each missing important feature, and train the predicted value as the real feature, so that the coverage of the feature can be improved; at the same time, we also combine important features to adaptively obtain augmented features as supplementary features, so as to improve the richness of the features. In terms of sample selection, the XGBOOST model is then learned to obtain a basic model, through which we can know which are difficult samples and which are possible noise samples. Through the scoring of the basic model, we find the difficult samples on the decision boundary, build a salvage model, and complement the poorly learned parts of the previous model; then remove the noise samples far from the decision boundary to obtain an anti-interference model, so that the interference of external conditions can be eliminated to a certain extent. Finally, the three models are integrated to obtain the final model score. The higher the model score, the higher the possibility of user demand for the financial product.

[0150] Further references Figure 3 , which shows an exemplary structural block diagram of an advertising and marketing modeling device based on the financial field according to an embodiment of the present application, which mainly includes:

[0151] The data acquisition unit 101 is used to acquire the collected user characteristic data as a sample set;

[0152] The sample set division unit 102 is used to divide the sample set into sample types according to the classification performance of the samples to obtain a plurality of sub-sample sets;

[0153] The multi-model extraction unit 103 is used to call the feature extraction model corresponding to each sample type, perform feature extraction on the corresponding sub-sample set, and fuse each set of feature values ​​generated by each feature extraction model;

[0154] The demand evaluation unit 104 is used to call the evaluation model to predict the demand for the financial product based on the fused feature value, and use the prediction result as the user's demand for the financial product.

[0155] It should be understood that the units described in the above device are similar to those in the reference Figure 1 The steps in the method described above correspond to each other. Therefore, the operations and features described above for the method are also applicable to the device and the units contained therein, and will not be repeated here. The device can be pre-implemented in the browser or other security application of the server, or loaded into the browser or its security application of the server by downloading or the like. The corresponding units in the device can cooperate with the units in the server to implement the solution of the embodiment of the present application.

[0156] For the several units mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units described above can be embodied in one unit. On the contrary, the features and functions of one unit described above can be further divided into multiple units to be embodied.

[0157] It should be noted that for details not disclosed in the advertising and marketing modeling device based on the financial field in the embodiment of the present application, please refer to the details disclosed in the above embodiments of the present application, which will not be repeated here.

[0158] Reference below Figure 4 , Figure 4 A schematic diagram of the structure of a computer system suitable for implementing a server of an embodiment of the present application is shown.

[0159] like Figure 4 As shown, the computer system includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 402 or the program loaded from the storage part 408 into the random access memory (RAM) 403. In the RAM 403, various programs and data required for the operation instructions of the system are also stored. The CPU 401, the ROM 402, and the RAM 403 are connected to each other through a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0160] The following components are connected to the I / O interface 405: an input section 406 including a keyboard, a mouse, etc.; an output section 407 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN card, a modem, etc. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the I / O interface 405 as needed. A removable medium 411, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 410 as needed, so that a computer program read therefrom is installed into the storage section 408 as needed.

[0161] It should be noted that the computer-readable medium shown in the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal can take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium such as a computer-readable storage medium that can send, propagate or transmit a program for use by or in conjunction with an instruction execution system, device or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0162] The flow chart and block diagram in the accompanying drawings illustrate the possible implementation architecture, function and operation instruction of the device and method according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of the code, and the aforementioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, the boxes represented by two connections can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation instruction, or can be implemented with a combination of dedicated hardware and computer instructions.

[0163] The units or modules involved in the embodiments described in the present application may be implemented by software or hardware. The units or modules described may also be arranged in a processor. The names of these units or modules do not, in some cases, constitute limitations on the units or modules themselves.

[0164] As another aspect, the present application further provides a computer-readable storage medium, which may be included in the server described in the above embodiment, or may exist independently without being assembled into the server. The above computer-readable storage medium stores one or more programs, and when the above programs are used by one or more processors to execute the advertising marketing modeling method based on the financial field described in the present application.

[0165] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the aforementioned disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other to form a technical solution.

Claims

1. An advertising marketing modeling method based on the financial field, characterized in that: include: Obtain the collected user feature data as a sample set; Dividing the sample set into sample types according to the classification performance of the samples to obtain a plurality of sub-sample sets; Calling the feature extraction model corresponding to each sample type, extracting features from the corresponding sub-sample set, and fusing the groups of feature values ​​generated by each of the feature extraction models; The evaluation model is called to predict the demand for the financial product based on the fused feature values, and the prediction result is used as the user's demand for the financial product.

2. The method according to claim 1, characterized in that The dividing the sample set into sample types according to the classification performance of the samples includes: Calling a pre-trained feature extraction model to perform sample evaluation on the sample set; Determine the dividing line between positive and negative samples in the evaluation results; Extracting a first proportion of samples from the sample set as a difficult sample set according to the distance from the segmentation line from near to far; According to the distance from the segmentation line from far to near, samples of a second proportion are extracted from the sample set as a noise sample set.

3. The method according to claim 2, characterized in that The calling of the feature extraction model corresponding to each sample type, performing feature extraction on the corresponding sub-sample set, and fusing each set of feature values ​​generated by each feature extraction model, includes: Calling the first feature extraction model to extract features from each sample in the difficult sample set, and using the generated features as the first features retrieved; Eliminating the noise sample set from the sample set, and calling the second feature extraction model to extract features from the remaining samples, and using the generated features as the second features after eliminating interference; Calling a third feature extraction model to extract features from each sample in the sample set, and using the generated features as the basic third features; Feature fusion is performed on the first feature, the second feature, and the third feature.

4. The method according to claim 1, characterized in that Before dividing the sample set into sample types according to the classification performance of the samples, the method further includes: Determine the missing core features in the user feature data as target features; Calling a pre-trained feature prediction model to predict missing data for the target feature, and backfilling the predicted data into the user feature data; The backfilled user feature data is used as the sample set.

5. The method according to claim 4, characterized in that Determine the missing core features in the user feature data, including: Input the user feature data into the LR model for evaluation to obtain the weight corresponding to each feature; Extract the first M user features in order according to the weights; Determine the coverage of each feature in the first M user features; The features with coverage rates higher than a threshold are removed from the first M user features, and the remaining features are used as the target features.

6. The method according to claim 1, characterized in that Before dividing the sample set into sample types according to the classification performance of the samples, the method further includes: Determining key features in the user feature data; Performing feature cross processing on the key features to generate new features; The user feature data and the new feature are combined as a sample set.

7. The method according to claim 6, characterized in that Determine the key features of the user characteristics, including: Input the user feature data into the LR model for evaluation to obtain the weight corresponding to each feature; Extract the first N user features in order according to the weights; Determine whether there is a gap in the weights of the first N user features; If there is a fault, all user features before the fault are used as the key features; If there is no fault, the first N user features are used as the key features.

8. An advertising and marketing modeling device based on the financial field, characterized in that: include: A data acquisition unit, used to acquire the collected user feature data as a sample set; A sample set division unit, used to divide the sample set into sample types according to the classification performance of the samples to obtain a plurality of sub-sample sets; A multi-model extraction unit, used to call the feature extraction model corresponding to each sample type, perform feature extraction on the corresponding sub-sample set, and fuse each set of feature values ​​generated by each of the feature extraction models; The demand assessment unit is used to call the assessment model to predict the demand for the financial product based on the fused feature value, and use the prediction result as the user's demand for the financial product.

9. A server comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images