Construction of business recommendation model, business recommendation method and device

Abstract

The application provides a kind of service recommendation model construction, service recommendation method and device, comprising: based on effective service features and derivative combination features, determine the data sample to be screened;And screen out positive samples and negative samples, generate modeling data;Utilize the initial service recommendation model of target service to predict modeling data, update the weight of negative sample according to the prediction result;Based on the updated weight of the negative sample, obtain the test data set from the modeling data, based on the test data set and the initial service recommendation model, construct the service recommendation model.The finally constituted construction service recommendation model fuses SVM model, random forest model, three-layer neural network model, XGBOOST model, and improves the generalization performance, and the constituted construction service recommendation model is used to predict the effective service features and derivative combination features of user, which can improve the success rate of service recommendation.

Description

Technical Field

[0001] This invention relates to the field of big data mining technology, and in particular to the construction of a business recommendation model, a business recommendation method, and an apparatus. Background Technology

[0002] The 5G Gold Coin service is a new service designed to enrich the customer experience and provide a reward service. Currently, there are two ways to recommend this service: the first is to select some high-value users and recommend the 5G Gold Coin service to them; the second is to determine whether to recommend the 5G Gold Coin service to users based on their historical acceptance of new plans.

[0003] However, the first recommendation method only considers the user's own value attributes without taking into account whether the user is willing to subscribe to the service, so the recommendation success rate is low. The second recommendation method does not take into account the differences between different packages for different users, and if a user subscribes to too many packages, it will also cause the user to have resistance, which also results in a low recommendation success rate. Summary of the Invention

[0004] This invention provides a business recommendation model construction, business recommendation method, and apparatus to address the low success rate of business recommendations in existing technologies and improve the success rate of business recommendations.

[0005] This invention provides a method for constructing a business recommendation model, comprising:

[0006] Based on the effective business features corresponding to the target business in the user data sample, obtain derived combination features;

[0007] Based on the effective business characteristics and the derived combination characteristics, the data samples to be screened are determined;

[0008] Positive and negative samples are selected from the data samples to be screened, and modeling data is generated by combining the positive and negative samples.

[0009] The modeling data is predicted using the initial business recommendation model of the target business, and the weights of the negative samples are updated based on the prediction results; wherein, the initial business recommendation model is constructed using SVM model, random forest model, three-layer neural network model, and XGBOOST model as a framework;

[0010] Based on the updated weights of the negative samples, the test dataset is obtained from the modeling data;

[0011] Based on the test dataset and the initial business recommendation model, the business recommendation model is constructed.

[0012] According to a method for constructing a business recommendation model provided by the present invention, the step of obtaining the effective business features includes:

[0013] Based on each business feature of the user data sample, a logistic regression model and a decision tree model are constructed, and a mutation calculation is performed based on the business features to obtain the logistic regression coefficient, the importance coefficient based on the decision tree, and the coefficient of variation for each business feature;

[0014] Based on the logistic regression coefficient, the importance coefficient based on the decision tree, and the coefficient of variation, the feature importance of each business feature is obtained;

[0015] Based on the importance of each of the business features, the effective business features of the user data sample are obtained.

[0016] According to a method for constructing a business recommendation model provided by the present invention, the step of obtaining derived combination features based on effective business features of user data samples includes:

[0017] A difference function is constructed based on every two of the aforementioned valid business features to obtain the difference weights;

[0018] Obtain the bucketing difference and ratio difference between the two valid business features;

[0019] Based on the difference weight, the binning difference, and the ratio difference, the derived combination features are obtained.

[0020] According to a method for constructing a business recommendation model provided by the present invention, the step of selecting positive and negative samples from the data samples to be screened, and combining the positive and negative samples to generate modeling data includes:

[0021] The data samples to be filtered are divided into positive samples and unknown samples; wherein, the positive samples are data samples corresponding to users who have subscribed to the target service, and the unknown samples are data samples corresponding to users who have not subscribed to the target service;

[0022] A subset of unknown samples is generated by sampling the unknown samples a predetermined number of times.

[0023] Based on the positive samples and each subset of the unknown samples, a classification model is constructed respectively;

[0024] Each of the classification models is used to predict a specified subset of unknown samples, and each classification model generates a prediction result set; wherein, the specified subset of unknown samples is the subset of unknown samples remaining other than the subset of unknown samples used to construct the classification model, and the prediction result set includes unknown samples in the specified subset of unknown samples whose probability of not subscribing to the target service is greater than a preset probability threshold.

[0025] Take the intersection of all the predicted result sets, and determine the unknown samples in the intersection as the negative samples;

[0026] The positive and negative samples are merged to generate the modeling data.

[0027] According to a method for constructing a business recommendation model provided by the present invention, before the step of merging the positive samples and the negative samples to generate the modeling data, the method includes:

[0028] Determine whether the proportion of the negative samples in the unknown samples reaches a preset proportion value. If not, resample the unknown samples to obtain more negative samples.

[0029] According to a method for constructing a business recommendation model provided by the present invention, the step of predicting the modeling data using an initial business recommendation model of the target business and updating the weights of the negative samples based on the prediction results includes:

[0030] Based on the modeling data, initialize the first weight of the positive sample and the second weight of the negative sample;

[0031] The initial business recommendation model is used to predict the modeling data to obtain a first probability value for the positive sample and a second probability value for the negative sample; wherein, the first probability value and the second probability value represent the probability of subscribing to the target business after recommending it to the user;

[0032] Obtain the similarity between each negative sample and all the positive samples;

[0033] The second weight is updated based on the similarity and the second probability value to obtain the updated second weight of the negative sample.

[0034] According to a method for constructing a business recommendation model provided by the present invention, the step of constructing the business recommendation model based on the test dataset and the initial business recommendation model includes:

[0035] Based on the frameworks of SVM model, random forest model, three-layer neural network model and XGBOOST model, the first model, the second model, the third model and the fourth model are constructed respectively;

[0036] Calculate the accuracy and model stability of the first model, the second model, the third model, and the fourth model on the test dataset, respectively.

[0037] The weights of the first model, the second model, the third model, and the fourth model are obtained based on the accuracy and model stability of the first model, the second model, the third model, and the fourth model, respectively.

[0038] The business recommendation model is constructed by fusing the weights of the first model, the second model, the third model, and the fourth model.

[0039] This invention also provides a business recommendation method, comprising:

[0040] Obtain the user's effective business characteristics and derived combined characteristics;

[0041] The effective business features and the derived combined features are input into the business recommendation model of the target business to obtain the output results;

[0042] Based on the output results, determine whether to recommend the target service to the user;

[0043] The business recommendation model is constructed using any of the above-described methods for constructing a business recommendation model.

[0044] The present invention also provides an apparatus for constructing a business recommendation model, comprising:

[0045] The first acquisition module is used to acquire derived combination features based on the effective business features corresponding to the target business in the user data sample;

[0046] The determination module is used to determine the data samples to be screened based on the effective business features and the derived combination features;

[0047] The generation module is used to filter out positive and negative samples from the data samples to be filtered, and combine the positive and negative samples to generate modeling data.

[0048] The second acquisition module is used to predict the modeling data using the initial business recommendation model of the target business, and update the weight of the negative sample according to the prediction result; wherein, the initial business recommendation model is constructed using SVM model, random forest model, three-layer neural network model, and XGBOOST model as a framework;

[0049] The third acquisition module is used to acquire the test dataset from the modeling data based on the weights updated by the negative samples;

[0050] A construction module is used to construct the business recommendation model based on the test dataset and the initial business recommendation model.

[0051] The present invention also provides a business recommendation device, comprising:

[0052] The first acquisition module is used to acquire the user's valid business characteristics and derived combination characteristics;

[0053] The second acquisition module is used to input the effective business features and the derived combination features into the business recommendation model of the target business and obtain the output results;

[0054] The service recommendation module is used to determine whether to recommend the target service to the user based on the output results;

[0055] The business recommendation model is constructed using any of the above-described methods for constructing a business recommendation model.

[0056] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for constructing a business recommendation model as described above, and / or the business recommendation method.

[0057] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for constructing a business recommendation model as described above, and / or the business recommendation method.

[0058] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a method for constructing a business recommendation model as described above, and / or the business recommendation method.

[0059] This invention provides a business recommendation model construction, method, and apparatus. By selecting features from user data samples, it provides effective business features and derived feature combinations, and mines true negative samples from the data samples to provide a data foundation for modeling. The initial business recommendation model predicts the modeling data, and updates the weights of negative samples based on the prediction results, thus addressing the potential for negative samples to be positive. This ensures that the updated weights better reflect the actual target business. Furthermore, the updated weights are used to filter the modeling data, obtaining a test dataset. Modeling is then performed based on the test dataset and the initial business recommendation model, which integrates four models: SVM, Random Forest, Three-Layer Neural Network, and XGBoost. This constructs a business recommendation model for the target business, improving its generalization performance. Using this constructed business recommendation model to predict effective business features and derived feature combinations for users increases the success rate of business recommendations. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating the method for constructing the business recommendation model provided by the present invention;

[0062] Figure 2 This is a flowchart illustrating the business recommendation method provided by the present invention;

[0063] Figure 3 This is a schematic diagram of the structure of the device for constructing the business recommendation model provided by the present invention;

[0064] Figure 4 This is a schematic diagram of the business recommendation device provided by the present invention;

[0065] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0067] It should be noted that in the description of the embodiments of the present invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The terms "upper," "lower," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly, for example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0068] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0069] Figure 1 This is a flowchart illustrating a method for constructing a business recommendation model provided by the present invention. Figure 1 As shown, the method includes:

[0070] Step 110: Based on the effective business features corresponding to the target business in the user data sample, obtain the derived combination features.

[0071] Step 120: Based on the effective business features and the derived combination features, determine the data samples to be screened.

[0072] Step 130: Select positive and negative samples from the data samples to be screened, and combine the positive and negative samples to generate modeling data.

[0073] Step 140: Use the initial business recommendation model of the target business to predict the modeling data, and update the weight of the negative sample based on the prediction result.

[0074] Step 150: Based on the updated weights of the negative samples, obtain the test dataset from the modeling data.

[0075] Step 160: Based on the test dataset and the initial business recommendation model, construct the business recommendation model.

[0076] The initial business recommendation model is constructed using SVM, random forest, three-layer neural network, and XGBOOST models as frameworks.

[0077] This invention provides a method for constructing a business recommendation model. By selecting features from user data samples, it provides effective business features and derived feature combinations, and mines real negative samples from the data samples to provide a data foundation for modeling. An initial business recommendation model is used to predict the modeling data, and the weights of negative samples are updated based on the prediction results. This addresses the potential for negative samples to be positive, making the updated weights more consistent with the actual target business. Modeling data is then filtered according to the updated weights to obtain a test dataset. Modeling is then performed based on the test dataset and the initial business recommendation model, which integrates four models: SVM, Random Forest, Three-Layer Neural Network, and XGBoost. This constructs a business recommendation model for the target business, improving the generalization performance of the business recommendation model. Using this constructed business recommendation model to predict effective business features and derived feature combinations for users can improve the success rate of business recommendations.

[0078] The following sections will describe steps 110-160 in detail.

[0079] Specifically, user data samples may include business characteristics and tags. Business characteristics may include user attributes (such as personal information such as gender, age, and education level), user's historical service plans, user's terminal (such as terminal type and model), and user consumption patterns.

[0080] The target service tags can be divided into two types: one is for users who have already subscribed to the target service, represented by 1; the other is for users who have not yet been recommended, so it is unknown whether they will subscribe to the service, represented by 0. In one embodiment, the target service can be a service offered by an operator. In another embodiment, the target service can be a 5G Coin service; the following description uses the 5G Coin service as an example.

[0081] In one embodiment, X can be usedi Let y represent the i-th business feature, such as user age, which can be represented by AGE = X1; and let y represent the label. Therefore, the user data sample can be expressed as:

[0082] sample=(X1,X2,…,X m ,y)

[0083] Where sample represents a user data sample, (X1,X2,…,X…) m Let ,y) represent business characteristics and tags. Therefore, the j-th user data sample can be expressed as:

[0084] sample j =(X 1j ,X 2j ,…,X mj ,y j )

[0085] In the target business, the original features may be redundant, which is not conducive to building a business recommendation model with good generalization ability. Therefore, feature selection is required to obtain features with good representation. In one embodiment, a feature selection algorithm based on index fluctuation is used to obtain the effective business features for each user data sample.

[0086] Specifically, since the 5G Gold Coin service doesn't have a truly negative class—that is, users who are recommended but don't subscribe to the 5G Gold Coin service—an unsupervised feature selection method is used. During the construction of the service recommendation model, it was found that features obtained by traditional unsupervised feature selection algorithms did not improve the model's generalization performance. Therefore, this invention proposes an indicator fluctuation-based feature selection algorithm based on 5G service practices to improve the model's recall. In practice, using the indicator fluctuation-based feature selection algorithm proposed in this invention, under the same conditions, the recall rate of the 5G Gold Coin service recommendation model can be improved by 2%.

[0087] In one embodiment, the steps for obtaining effective business features using an indicator fluctuation feature selection algorithm include: constructing a logistic regression model and a decision tree model based on each business feature of the user data sample; performing mutation calculation based on the business features to obtain the logistic regression coefficient, importance coefficient based on the decision tree, and coefficient of variation for each business feature; obtaining the feature importance of each business feature based on the logistic regression coefficient, importance coefficient based on the decision tree, and coefficient of variation; and obtaining the effective business features of the user data sample based on the feature importance of each business feature.

[0088] Each sub-step is described in detail below:

[0089] 1. Construct logistic regression and decision tree models based on business characteristics, and perform mutation calculations based on business characteristics to obtain the logistic regression coefficient, importance coefficient, and coefficient of variation for each business characteristic.

[0090] (1) Construct a logistic regression model based on business characteristics and obtain the logistic regression coefficients for each business characteristic.

[0091] For ease of description, we will use X1, X2, ..., X m1 Let m1 represent the number of business features. Based on these business features, we first assume that unlabeled users are negative users, construct a logistic regression model, and obtain the logistic regression coefficients for each feature. i = 1, 2, ..., m1 represents the logistic regression coefficient. Since each logistic regression coefficient is based on a real number from negative infinity to positive infinity, the parameters need to be normalized. The normalization formula for the logistic regression coefficients is:

[0092]

[0093] (2) Construct a decision tree model based on business characteristics and obtain the importance coefficient of each business characteristic based on the decision tree. The importance coefficient is used... i = 1, 2, ..., m1 represents the number of elements in a decision tree. The normalization formula for the importance coefficient based on the decision tree is:

[0094]

[0095] (3) Perform mutation calculations based on business characteristics to obtain the coefficient of variation for each characteristic. The coefficient of variation is used... i = 1, 2, ..., m1 represents the expression. The formula for calculating the coefficient of variation is:

[0096]

[0097] Where, x ij Represents business characteristic X i The j-th sample instance, where n represents the number of user data samples, indicating... The coefficient of variation of the business characteristics is normalized using the following formula:

[0098]

[0099] 2. Based on logistic regression coefficients, importance coefficients based on decision trees, and coefficients of variation, obtain the feature importance of each business feature.

[0100] (1) The three coefficients of the business features are fused. Specifically, linear and nonlinear fusion methods are used to obtain the preliminary fused feature importance. The formula for expressing the preliminary fused feature importance is as follows:

[0101]

[0102] in, A linear fusion representing the importance of features is used, assigning a weight of 0.2 to each metric. Linear fusion representing the importance of features is assigned a weight of 0.4.

[0103] (2) Calculation as well as Each is ranked according to its own total importance; the ranking can be used as follows: as well as This means calculating the standard deviation of the three rankings, using ss i express.

[0104] (3) Combine the standard deviation with the preliminary fusion feature importance to obtain the final feature importance. The calculation formula is as follows:

[0105]

[0106] in, For the final feature importance, m represents the number of features.

[0107] 3. Based on the feature importance of each business feature, obtain the effective business features from the user data sample. Specifically, based on the feature importance calculated for each business feature, sort the business features (e.g., sort from largest to smallest), and select a preset proportion of business features as effective business features. For example, select the top 70% of business features as effective business features. It is important to understand that the ultimately obtained effective business features are the modeling features for building the business recommendation model.

[0108] It should be understood that the specific weights assigned to the logistic regression coefficients, decision tree-based importance coefficients, and coefficients of variation are based on the 5G Gold Coin service. Different target services may be assigned different weights. Therefore, the feature importance of each feature corresponding to different target services may also be different, resulting in different effective service features being selected.

[0109] In one embodiment, a bucket-based feature combination algorithm is used to obtain derived combined features based on effective business features.

[0110] Specifically, traditional linear feature combination requires sales personnel to have extensive business experience and the ability to keenly identify relationships between business features in order to combine them. This invention, based on business practice, proposes a bucketed fusion feature combination algorithm that can combine effective features even when sales personnel are not particularly familiar with the business. In practice, the recall rate of the 5G Gold Coin service recommendation model can be improved by 1.33%.

[0111] In one embodiment, a bucketed feature combination algorithm can be used to obtain derived combined features based on effective business features. Specifically, the step of obtaining derived combined features based on the effective business features of user data samples includes: constructing a difference function based on every two effective business features to obtain difference weights; obtaining the bucketed difference and ratio difference between two effective business features; and obtaining the derived combined features based on the difference weights, the bucketed difference, and the ratio difference.

[0112] Each sub-step is described in detail below:

[0113] 1. Construct a difference function based on every two of the aforementioned valid business features, and obtain the difference weights.

[0114] (1) Select two valid business features, namely X1 = (x 11 ,x 12 ,…,x 1n ) and X2 = (x 21 ,x 22 ,…,x 2n Each valid business feature is divided into 10 bins in ascending order. The ten bins of X1 are indicated by... The tenth symbol of X2 is

[0115] (2) Calculate the mean values ​​of effective business features X1 and X2, respectively using... and express.

[0116] (3) Construct the difference function, the specific expression of which is as follows:

[0117]

[0118] Represented as x 1j and x 2j The ratio of the summation to the product, multiplied by 0.5 after taking the logarithm, can be used to represent the difference between linear and nonlinear operations, and can be used as the first-level difference weight.

[0119] 2. Obtain the bucketing difference and ratio difference between the two valid business features.

[0120] Assume x 1j In the bucket In the middle, x 2j In the bucket In the diagram, x is represented by j1 and j2. 1j and x 2j Therefore, calculating the difference between the two from the perspectives of binning and the mean, the binning difference is:

[0121]

[0122] Calculate x 1j and x 2j The ratio difference is:

[0123]

[0124] 3. Based on the difference weights, the binning differences, and the ratio differences, obtain the derived combined features. Specifically, the difference function value is used as the weight (difference weight), and the binning differences and ratio differences are used as linear and non-linear feature derivation representations to constitute the final derived combined features. The specific expression is as follows:

[0125]

[0126] Where C(X1,X2) represents the derived combined features. 0.65 represents the proportion of binning differences in the derived combined features, and 0.35 represents the proportion of ratio differences in the derived combined features. It should be understood that the proportions of 0.65 and 0.35 in the above formula are based on the proportions set for the 5G Gold Coin business, and the proportions may be different for different target businesses.

[0127] By combining two effective business features in pairs, multiple derived combination features can be obtained. Based on the calculation size of the derived combination features, some derived combination features can be selected and combined with the effective business features to form a data sample to be screened.

[0128] In one embodiment, a PU probabilistic iterative negative sample mining algorithm can be used to mine true negative samples. It's important to understand that only positive samples (users who have subscribed to the target service) can be identified in the data samples to be screened; the rest are unknown samples, representing users who have not subscribed to the target service. Since unknown samples can contain both positive and negative samples, building a classification model based solely on positive and unknown samples is unrealistic. To accurately recommend the target service to customers, it's necessary to mine true negative samples—those that explicitly reject recommendations from users who haven't subscribed to the target service. This invention proposes a PU probabilistic iterative negative sample mining algorithm that can effectively mine negative samples.

[0129] Specifically, the step of selecting positive and negative samples from the data samples to be screened, and generating modeling data by combining the positive and negative samples, includes: dividing the data samples to be screened into positive samples and unknown samples; wherein, the positive samples are data samples corresponding to users who have subscribed to the target service, and the unknown samples are data samples corresponding to users who have not subscribed to the target service; sampling the unknown samples a preset number of times to generate the unknown sample subset; constructing a classification model based on the positive samples and each unknown sample subset; using each classification model to predict a specified unknown sample subset, and generating a prediction result set for each classification model; wherein, the specified unknown sample subset is the remaining unknown sample subset excluding the unknown sample subset used to construct the classification model, and the prediction result set includes unknown samples in the specified unknown sample subset whose probability of not subscribing to the target service is greater than a preset probability threshold; taking the intersection of all prediction result sets, and determining the unknown samples in the intersection as the negative samples; merging the positive samples and negative samples to generate the modeling data.

[0130] Each sub-step is described in detail below:

[0131] 1. The data samples to be screened are divided into positive samples and unknown samples; wherein, the positive samples are data samples corresponding to users who have subscribed to the target service, and the unknown samples are data samples corresponding to users who have not subscribed to the target service. Specifically, P represents all positive samples, and U represents all unknown samples.

[0132] 2. Sampling is performed a predetermined number of times from the unknown samples to generate a subset of the unknown samples. i = 1, 2, ..., 15. For example, sampling 15 times from an unknown sample generates a subset of unknown samples.

[0133] 3. Based on the positive samples and each of the unknown sample subsets, construct classification models respectively. Specifically, merge P and one of the unknown sample subsets respectively. Generate samples Assume the unknown samples are negative samples and construct a classification model. i It should be understood that 15 classification models can be built.

[0134] The classification model constructed in this step can be built using basic classification models and classifiers as a framework.

[0135] 4. Each of the classification models is used to predict a specified subset of unknown samples, and each classification model generates a prediction result set; wherein, the specified subset of unknown samples is the subset of unknown samples remaining other than the subset of unknown samples used to construct the classification model, and the prediction result set includes unknown samples in the specified subset of unknown samples whose probability of not subscribing to the target service is greater than a preset probability threshold.

[0136] Specifically, using Model i right For j∈{1,2,…,15} and j≠i, a prediction is made to obtain a probability value of 0, where 0 represents a user who does not subscribe to the target service. A set of users with a probability value greater than a preset probability threshold is also obtained. i→j For Set ,j∈{1,2,…,15} andj≠i, i→j Taking the union is equivalent to taking the model as the set of its constituent elements. i The result that the probability of the remaining 14 unknown sample subsets being predicted as 0 is greater than the preset probability threshold is used. This indicates that, in one example, the preset probability threshold is 0.8.

[0137] It should be understood that each classification model needs to make predictions for the remaining 14 unknown sample subsets, and all classification models generate a total of 15 prediction result sets. i = 1, 2, ..., 15.

[0138] 5. Take the intersection of all the predicted result sets, and determine the unknown samples in the intersection as the negative samples; merge the positive samples and the negative samples to generate the modeling data. Specifically, for 15... Take the intersection to generate the required true negative samples, using N True express.

[0139] In one embodiment, before the step of merging the positive samples and the negative samples to generate the modeling data, the method includes: determining whether the proportion of the negative samples in the unknown samples reaches a preset proportion value; if not, re-sampling the unknown samples to obtain more negative samples.

[0140] Specifically, by setting a preset ratio, it is determined whether the proportion of generated negative samples in the unknown samples reaches the preset ratio, such as 10%, or in other words, whether the number of unknown samples has decreased by 10%. If it does not reach 10%, N can be... True Add to the list of true negative samples, then remove N from the unknown samples. True Continue iterating through sub-steps 1-5 until 10% is achieved; if 10% is reached, merge the positive and negative samples to generate modeling data.

[0141] Traditional classification models have proposed numerous sample weight update strategies. However, in many target services (such as 5G coin services), the effectiveness of mined negative samples cannot be fully guaranteed. Dynamically updating sample weights can increase the amount of erroneous sample information, thus requiring new sample weight update strategies. In one embodiment, a dynamic negative sample weight update algorithm can be used to update the weights of negative samples, thereby reducing the amplification of erroneous information in negative samples. In practice, using the dynamic negative sample weight update algorithm proposed in this invention, the recall rate of the business recommendation model improved by 0.52%.

[0142] Specifically, the step of using the initial business recommendation model of the target business to predict the modeling data and updating the weights of the negative samples based on the prediction results includes: initializing a first weight of the positive samples and a second weight of the negative samples based on the modeling data; using the initial business recommendation model to predict the modeling data and obtaining a first probability value of the positive samples and a second probability value of the negative samples; wherein the first probability value and the second probability value represent the probability of subscribing to the target business after recommending it to the user; obtaining the similarity between each negative sample and all the positive samples; updating the second weights based on the similarity and the second probability values ​​to obtain the updated second weights of the negative samples.

[0143] Each sub-step is described in detail below:

[0144] 1. Based on the modeling data, initialize the first weight of the positive samples and the second weight of the negative samples. The first weight of the positive samples is... i = 1, 2, ..., n1 represents the second weight of the negative sample. j = 1, 2, ..., n2 represents

[0145] 2. Use the initial business recommendation model to predict the modeling data and obtain the first probability value of the positive sample and the second probability value of the negative sample; wherein, the first probability value and the second probability value represent the probability of subscribing to the target business after recommending it to the user.

[0146] Specifically, based on the first and second weights initialized in step 1, an initial business recommendation model is trained. Predictions are then made on the modeling data to obtain a probability value of 1, which represents the probability that the user will subscribe to the target user after a recommendation is made. The probability value obtained from positive sample predictions is used... i = 1, 2, ..., n1 represents the probability value obtained from the prediction of negative samples, which is denoted by ... j = 1, 2, ..., n2 represents this.

[0147] 3. Obtain the similarity between each negative sample and all positive samples. Specifically, calculate the mean similarity between a negative sample and all positive samples using the following formula:

[0148]

[0149] Among them, smi j This represents the mean distance between negative sample j and all positive samples. This represents the absolute error between negative sample j and positive sample i. This represents the actual feature value of the positive sample. This represents the actual feature value of the negative sample.

[0150] 4. Update the second weight based on the similarity and the second probability value to obtain the updated second weight for the negative sample. Specifically, the updated second weight is:

[0151]

[0152] Where α is the learning rate, and in one example, the learning rate α is set to 0.001. According to the above formula, the update strategy for negative sample weights mainly consists of two parts: the first part is... The higher the probability of a negative sample being 1, the greater the deviation in the prediction for that sample. Therefore, it needs to be given more attention in the next modeling iteration, and its weight should be increased. The weights are removed from the iterated samples, i.e., compared with the left side of the formula. Proportional; considering that negative samples are mined by algorithms, there is a possibility that one of them is itself a positive sample. Therefore, it is necessary to calculate the similarity between this sample and the true positive samples, using SMI. j It means, smi j A larger value indicates that the sample is more likely to be a positive sample, so the weight of negative samples needs to be reduced, i.e., SMI. j The update of negative sample weights is inversely proportional to the update of negative sample weights. To reflect this, (1-smi) j This is introduced into the iterative formula, which is the second part of the update strategy.

[0153] The second weight of the negative samples is updated, and the first weight of the positive samples is also updated accordingly. Based on the updated weights of the positive and negative samples, the test dataset is selected from the modeling data. The weights of the positive and negative samples in the test dataset are the updated first and second weights.

[0154] In the process of building the 5G Gold Coin business recommendation model, it was found that simple model integration could not efficiently improve the model's recall rate. The analysis revealed that this was related to the model's accuracy and stability.

[0155] To improve the recall rate of the model, in one embodiment, a business recommendation model can be constructed using the Bagging-Stable model ensemble framework. Specifically, the steps of constructing the business recommendation model based on the test dataset and the initial business recommendation model include: constructing a first model, a second model, a third model, and a fourth model based on SVM, Random Forest, a three-layer neural network, and XGBOOST models, respectively; calculating the accuracy and model stability of the first model, the second model, the third model, and the fourth model on the test dataset; obtaining the weights of the first model, the second model, the third model, and the fourth model based on the accuracy and model stability of the first model, the second model, the third model, and the fourth model, respectively; and fusing the weights of the first model, the second model, the third model, and the fourth model to construct the business recommendation model.

[0156] Each sub-step is described in detail below:

[0157] 1. Based on the SVM model, random forest model, three-layer neural network model, and XGBoost model as frameworks, construct the first model, the second model, the third model, and the fourth model, respectively. Specifically, they are represented by Model1, Model2, Model3, and Model4.

[0158] 2. Calculate the accuracy and model stability of the first model, the second model, the third model, and the fourth model on the test dataset, respectively.

[0159] (1) Calculate Model i The recall r on the test dataset i The softmax function is used to calculate the recall probability of each model, which represents the model's accuracy. The formula for calculating the model's accuracy is as follows:

[0160]

[0161] (2) Calculate Model i The error between the predicted probability of a positive sample and the true label, where the number of positive samples is n1, and the probability of a single predicted sample j is... The superscript i represents the model, and the subscript j represents the sample. This represents the error of the positive samples. In practice, the error is corrected by dividing by the mean error. The standard deviation of the correction error represents the model's stability. The formula for calculating model stability is as follows:

[0162]

[0163] 3. Based on the accuracy and model stability of the first model, the second model, the third model, and the fourth model, respectively, obtain the weights of the first model, the second model, the third model, and the fourth model.

[0164] use The accuracy of the model is represented by estd. i To represent model stability, the model parameter function is constructed, and the specific formula is as follows:

[0165]

[0166] in, It represents a fusion of nonlinear accuracy and stability. It represents linear fusion. The weights of the stable models are represented by ratios, thus biasing the overall system towards the more stable models. This results in the final weights of the four models.

[0167] 4. Based on the weights of the first model, the second model, the third model, and the fourth model, the business recommendation model is constructed by fusing them. Specifically, the four models are fused to obtain the final business recommendation model, as shown in the following formula:

[0168]

[0169]

[0170] In practice, the 5G Gold Coin service recommendation model constructed using the above steps improved the recall rate by 1.15%.

[0171] In summary, the method for constructing a business recommendation model provided by this invention has the following technical effects:

[0172] (1) Based on actual business scenarios, an indicator fluctuation feature selection algorithm is proposed. Combining logistic regression model (linear model), decision tree model and information of the coefficient of variation, linear and nonlinear information are integrated to construct indicator fluctuation feature selection, thereby more effectively selecting features, obtaining effective business features, and improving the recall rate of the model.

[0173] (2) Based on the bucketing fusion feature combination algorithm, derived combined features are obtained. A difference function is constructed as the weight, and then the bucketing and ratio information are fused to construct a new feature derivation system, which automatically generates effective derived combined features, reducing the workload of business personnel and improving the recall rate of the model.

[0174] (3) Based on the PU probabilistic iterative negative sample mining algorithm, real negative samples are mined. Through multiple sampling, the effective negative samples are obtained by first taking the union and then the intersection. By setting the iteration stopping condition, modelable negative samples are finally obtained, providing a data foundation for modeling.

[0175] (4) The negative sample weight dynamic update algorithm updates the weight of negative samples by fusing information on the similarity between negative and positive samples and their own predicted probability values. The negative sample weight update process not only solves the hidden danger that the sample may be positive, but also dynamically updates the weight based on its own prediction results. The weight update is in line with the actual situation of 5G Gold Coin business.

[0176] (5) By integrating the accuracy and stability of multiple models and then assigning different weights to different models through function transformation, the generalization performance of the models is improved by linearly weighting these models.

[0177] like Figure 2 As shown, the present invention also provides a business recommendation method, comprising:

[0178] Step 210: Obtain the user's valid business characteristics and derived combination characteristics;

[0179] Step 220: Input the effective business features and the derived combination features into the business recommendation model of the target business, and obtain the output results;

[0180] Step 230: Determine whether to recommend the target service to the user based on the output result;

[0181] The business recommendation model is constructed using any of the above-mentioned methods for constructing a business recommendation model.

[0182] The following describes the apparatus for constructing the business recommendation model and the business recommendation apparatus provided by the present invention. The apparatus for constructing the business recommendation model described below can be referred to in correspondence with the method for constructing the business recommendation model and the business recommendation method described above.

[0183] like Figure 3 As shown, the present invention also provides an apparatus for constructing a business recommendation model, comprising:

[0184] The first acquisition module 310 is used to acquire derived combination features based on the effective business features corresponding to the target business in the user data sample;

[0185] The determination module 320 is used to determine the data samples to be screened based on the effective business features and the derived combination features;

[0186] The generation module 330 is used to filter out positive and negative samples from the data samples to be filtered, and combine the positive and negative samples to generate modeling data.

[0187] The second acquisition module 340 is used to predict the modeling data using the initial business recommendation model of the target business, and update the weight of the negative sample according to the prediction result; wherein, the initial business recommendation model is constructed using SVM model, random forest model, three-layer neural network model, and XGBOOST model as a framework;

[0188] The third acquisition module 350 is used to acquire a test dataset from the modeling data based on the weights updated by the negative samples.

[0189] The construction module 360 ​​is used to construct the business recommendation model based on the test dataset and the initial business recommendation model.

[0190] like Figure 4 As shown, the present invention also provides a business recommendation device, comprising:

[0191] The first acquisition module 410 is used to acquire the user's valid business characteristics and derived combination characteristics;

[0192] The second acquisition module 420 is used to input the effective business features and the derived combination features into the business recommendation model of the target business and obtain the output results;

[0193] The service recommendation module 430 is used to determine whether to recommend the target service to the user based on the output result.

[0194] The business recommendation model is constructed using any of the above-described methods for constructing a business recommendation model.

[0195] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a method for constructing a business recommendation model. This method includes: obtaining derived combined features based on effective business features corresponding to a target business in user data samples; determining data samples to be screened based on the effective business features and the derived combined features; selecting positive and negative samples from the data samples to be screened, and generating modeling data by combining the positive and negative samples; predicting the modeling data using an initial business recommendation model for the target business, and updating the weights of the negative samples based on the prediction results; obtaining a test dataset from the modeling data based on the updated weights of the negative samples; and constructing the business recommendation model based on the test dataset and the initial business recommendation model. The initial business recommendation model is constructed using an SVM model, a random forest model, a three-layer neural network model, and an XGBOOST model as a framework.

[0196] And / or, perform a business recommendation method, the method comprising: obtaining effective business features and derived combination features of a user; inputting the effective business features and the derived combination features into a business recommendation model of a target business, and obtaining an output result; determining whether to recommend the target business to the user based on the output result; wherein the business recommendation model is constructed using any of the above-described methods for constructing a business recommendation model.

[0197] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0198] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the business recommendation model construction method provided by the above methods. The method includes: obtaining derived combination features based on effective business features corresponding to the target business in user data samples; determining data samples to be screened based on the effective business features and the derived combination features; selecting positive and negative samples from the data samples to be screened, and generating modeling data by combining the positive and negative samples; predicting the modeling data using the initial business recommendation model of the target business, and updating the weights of the negative samples according to the prediction results; obtaining a test dataset from the modeling data based on the updated weights of the negative samples; and constructing the business recommendation model based on the test dataset and the initial business recommendation model. The initial business recommendation model is constructed using an SVM model, a random forest model, a three-layer neural network model, and an XGBOOST model as a framework.

[0199] And / or, capable of executing a business recommendation method, the method comprising: acquiring effective business features and derived combination features of a user; inputting the effective business features and the derived combination features into a business recommendation model of a target business, and acquiring an output result; determining whether to recommend the target business to the user based on the output result; wherein, the business recommendation model is constructed using any of the above-described methods for constructing a business recommendation model.

[0200] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for constructing a business recommendation model provided by the above methods. This method includes: obtaining derived combination features based on effective business features corresponding to a target business in user data samples; determining data samples to be screened based on the effective business features and the derived combination features; selecting positive and negative samples from the data samples to be screened, and generating modeling data by combining the positive and negative samples; predicting the modeling data using an initial business recommendation model for the target business, and updating the weights of the negative samples based on the prediction results; obtaining a test dataset from the modeling data based on the updated weights of the negative samples; and constructing the business recommendation model based on the test dataset and the initial business recommendation model. The initial business recommendation model is constructed using an SVM model, a random forest model, a three-layer neural network model, or an XGBOOST model as a framework.

[0201] And / or, perform a business recommendation method, the method comprising: obtaining effective business features and derived combination features of a user; inputting the effective business features and the derived combination features into a business recommendation model of a target business, and obtaining an output result; determining whether to recommend the target business to the user based on the output result; wherein the business recommendation model is constructed using any of the above-described methods for constructing a business recommendation model.

[0202] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0203] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a business recommendation model, characterized in that, include: Based on the effective business features corresponding to the target business in the user data sample, obtain derived combination features; Based on the effective business characteristics and the derived combination characteristics, the data samples to be screened are determined; Positive and negative samples are selected from the data samples to be screened, and modeling data is generated by combining the positive and negative samples. The modeling data is predicted using the initial business recommendation model of the target business, and the weights of the negative samples are updated based on the prediction results. Based on the updated weights of the negative samples, the test dataset is obtained from the modeling data; Based on the test dataset and the initial business recommendation model, the business recommendation model is constructed; The initial business recommendation model is constructed using SVM model, random forest model, three-layer neural network model, and XGBOOST model as a framework. The steps for constructing the business recommendation model based on the test dataset and the initial business recommendation model include: Based on the frameworks of SVM model, random forest model, three-layer neural network model and XGBOOST model, the first model, the second model, the third model and the fourth model are constructed respectively; Calculate the accuracy and model stability of the first model, the second model, the third model, and the fourth model on the test dataset, respectively. The weights of the first model, the second model, the third model, and the fourth model are obtained based on the accuracy and model stability of the first model, the second model, the third model, and the fourth model, respectively. The business recommendation model is constructed by fusing the weights of the first model, the second model, the third model, and the fourth model.

2. The method for constructing a business recommendation model according to claim 1, characterized in that, The steps for obtaining the valid business characteristics include: Based on each business feature of the user data sample, a logistic regression model and a decision tree model are constructed, and a mutation calculation is performed based on the business features to obtain the logistic regression coefficient, the importance coefficient based on the decision tree, and the coefficient of variation for each business feature; Based on the logistic regression coefficient, the importance coefficient based on the decision tree, and the coefficient of variation, the feature importance of each business feature is obtained; Based on the importance of each of the business features, the effective business features of the user data sample are obtained.

3. The method for constructing a business recommendation model according to claim 1, characterized in that, The steps for obtaining derived combined features based on effective business features of user data samples include: A difference function is constructed based on every two of the aforementioned valid business features to obtain the difference weights; Obtain the bucketing difference and ratio difference between the two valid business features; Based on the difference weight, the binning difference, and the ratio difference, the derived combination features are obtained.

4. The method for constructing a business recommendation model according to claim 1, characterized in that, The step of selecting positive and negative samples from the data samples to be screened, and combining the positive and negative samples to generate modeling data includes: The data samples to be filtered are divided into positive samples and unknown samples; wherein, the positive samples are data samples corresponding to users who have subscribed to the target service, and the unknown samples are data samples corresponding to users who have not subscribed to the target service; A subset of unknown samples is generated by sampling the unknown samples a predetermined number of times. Based on the positive samples and each subset of the unknown samples, a classification model is constructed respectively; Each of the classification models is used to predict a specified subset of unknown samples, and each classification model generates a prediction result set; wherein, the specified subset of unknown samples is the subset of unknown samples remaining other than the subset of unknown samples used to construct the classification model, and the prediction result set includes unknown samples in the specified subset of unknown samples whose probability of not subscribing to the target service is greater than a preset probability threshold. Take the intersection of all the predicted result sets, and determine the unknown samples in the intersection as the negative samples; The positive and negative samples are merged to generate the modeling data.

5. The method for constructing a business recommendation model according to claim 4, characterized in that, Before the step of merging the positive samples and the negative samples to generate the modeling data, the following steps are included: Determine whether the proportion of the negative samples in the unknown samples reaches a preset proportion value. If not, resample the unknown samples to obtain more negative samples.

6. The method for constructing a business recommendation model according to claim 1, characterized in that, The step of using the initial business recommendation model of the target business to predict the modeling data and updating the weights of the negative samples based on the prediction results includes: Based on the modeling data, initialize the first weight of the positive sample and the second weight of the negative sample; The initial business recommendation model is used to predict the modeling data to obtain a first probability value for the positive sample and a second probability value for the negative sample; wherein, the first probability value and the second probability value represent the probability of subscribing to the target business after recommending it to the user; Obtain the similarity between each negative sample and all the positive samples; The second weight is updated based on the similarity and the second probability value to obtain the updated second weight of the negative sample.

7. A business recommendation method, characterized in that, include: Obtain the user's effective business characteristics and derived combined characteristics; The effective business features and the derived combined features are input into the business recommendation model of the target business to obtain the output results; Based on the output results, determine whether to recommend the target service to the user; The business recommendation model is constructed using the construction method of the business recommendation model according to any one of claims 1 to 6.

8. An apparatus for constructing a business recommendation model, characterized in that, include: The first acquisition module is used to acquire derived combination features based on the effective business features corresponding to the target business in the user data sample; The determination module is used to determine the data samples to be screened based on the effective business features and the derived combination features; The generation module is used to filter out positive and negative samples from the data samples to be filtered, and combine the positive and negative samples to generate modeling data. The second acquisition module is used to predict the modeling data using the initial business recommendation model of the target business, and update the weight of the negative sample according to the prediction result; wherein, the initial business recommendation model is constructed using SVM model, random forest model, three-layer neural network model, and XGBOOST model as a framework; The third acquisition module is used to acquire the test dataset from the modeling data based on the weights updated by the negative samples; A construction module is used to construct the business recommendation model based on the test dataset and the initial business recommendation model; The construction module is further configured to construct a first model, a second model, a third model, and a fourth model based on SVM, random forest, three-layer neural network, and XGBOOST models, respectively; calculate the accuracy and model stability of the first model, the second model, the third model, and the fourth model on the test dataset; obtain the weights of the first model, the second model, the third model, and the fourth model based on the accuracy and model stability of the first model, the second model, the third model, and the fourth model, respectively; and fuse the weights of the first model, the second model, the third model, and the fourth model to construct the business recommendation model.

9. A business recommendation device, characterized in that, include: The first acquisition module is used to acquire the user's valid business characteristics and derived combination characteristics; The second acquisition module is used to input the effective business features and the derived combination features into the business recommendation model of the target business and obtain the output results; The service recommendation module is used to determine whether to recommend the target service to the user based on the output results; The business recommendation model is constructed using the construction method of the business recommendation model according to any one of claims 1 to 6.

Images