Channel success rate prediction method and device, electronic equipment and computer program product
By employing cascaded data elimination, sample expansion, and feature fusion filtering, combined with probability residual jump connection method to optimize the loss function, the problem of low accuracy in channel success rate prediction was solved, thereby improving the accuracy of channel success rate prediction and the model performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA MOBILE GROUP JIANGSU
- Filing Date
- 2022-03-14
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for predicting channel success rates are inaccurate, market research is costly and it is difficult to obtain real user feedback, while historical experience methods are easily affected by policy changes.
By removing outliers through data cascading and filtering, expanding the sample and fusing features, a channel success rate prediction model is constructed, and the loss function is optimized using a probabilistic residual jump connection method.
It improved the accuracy of channel success rate prediction and enhanced the model's generalization ability and prediction performance.
Smart Images
Figure CN116796180B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, specifically to a method, apparatus, electronic device, and computer program product for predicting channel success rates. Background Technology
[0002] Currently, there are two main methods for predicting channel success rates. The first is to determine the current channel's success rate by analyzing user feedback through market research. The second is to determine the current channel's success rate by analyzing historical channel success rates.
[0003] However, the first method mentioned above requires market research, which is costly, and it's difficult to guarantee obtaining genuine user feedback. The second method, on the other hand, is susceptible to policy changes. Therefore, the accuracy of channel success rate predictions based on existing forecasting methods is low. Summary of the Invention
[0004] This application provides a channel success rate prediction method, apparatus, electronic device, and computer program product to solve the technical problem of low accuracy in current channel success rate prediction.
[0005] In a first aspect, embodiments of this application provide a method for predicting channel success rates, including:
[0006] Based on the initial sample data, cascaded data elimination is performed to obtain the first sample data;
[0007] Based on the first sample data, sample expansion is performed to obtain the second sample data;
[0008] Based on the first sample data and the second sample data, feature fusion and feature filtering are performed to obtain the target features;
[0009] Based on the target features, channel success rate is predicted to obtain an initial channel success rate, and a target loss function is determined based on the initial channel success rate.
[0010] A channel success rate prediction model is constructed based on the target loss function and the probability residual jump connection method. The channel success rate is then predicted based on the channel success rate prediction model and the channel data to be predicted, and the target channel success rate is obtained.
[0011] In one embodiment, the step of performing cascaded data elimination based on initial sample data to obtain the first sample data includes:
[0012] Construct a first joint distribution based on the single feature variable and the target variable in the initial sample data;
[0013] The target anomaly data is determined based on the probability density of the first joint distribution;
[0014] The target abnormal data is removed from the initial sample data to obtain the first sample data.
[0015] In one embodiment, the step of determining the target anomaly data based on the probability density of the first joint distribution includes:
[0016] The initial outlier data is determined based on the probability density of the first joint distribution;
[0017] Construct a second joint distribution of all feature variables and the target variable in the initial abnormal data;
[0018] The target anomaly data is determined based on the probability density of the second joint distribution.
[0019] In one embodiment, the step of determining the target anomaly data based on the probability density of the second joint distribution includes:
[0020] Calculate the probability density of the second joint distribution;
[0021] The probability density of the second joint distribution is compared with a preset threshold;
[0022] If the probability density of the second joint distribution is less than the preset threshold, then all feature variables in the second joint distribution are identified as target abnormal data.
[0023] In one embodiment, the step of augmenting the sample data based on the first sample data to obtain the second sample data includes:
[0024] Multiple intermediate sample data are generated from the first sample data using multiple initial generators;
[0025] The summation of multiple intermediate sample data is then input into the discriminator to obtain the predicted value;
[0026] Calculate the loss value based on the predicted value;
[0027] Based on the loss value, the loss function of each initial generator is corrected to obtain multiple target generators;
[0028] The first sample data is augmented using multiple target generators to obtain the second sample data.
[0029] In one embodiment, the step of performing feature fusion and feature filtering based on the first sample data and the second sample data to obtain the target features includes:
[0030] The first sample data and the second sample data are fused to obtain fused features;
[0031] The fused features are subjected to feature dimensionality reduction and filtering based on preset regularization to obtain the target features.
[0032] In one embodiment, the step of determining the target loss function based on the initial channel success rate includes:
[0033] The error percentage is determined based on the initial channel success rate and the actual value corresponding to the target feature;
[0034] Determine the ratio of the initial channel success rate to the actual value corresponding to the target feature;
[0035] The target loss function is determined based on the error percentage and the ratio.
[0036] Secondly, embodiments of this application provide a channel success rate prediction device, comprising:
[0037] The elimination module is used to perform cascaded elimination of data based on the initial sample data to obtain the first sample data;
[0038] An expansion module is used to expand the sample data based on the first sample data to obtain the second sample data.
[0039] The feature processing module is used to perform feature fusion and feature filtering based on the first sample data and the second sample data to obtain target features;
[0040] The determination module is used to predict the channel success rate based on the target features, obtain an initial channel success rate, and determine the target loss function based on the initial channel success rate.
[0041] The module is used to construct a channel success rate prediction model based on the target loss function and the probability residual jump connection method, and to predict the channel success rate based on the channel success rate prediction model and the channel data to be predicted, and obtain the target channel success rate.
[0042] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the channel success rate prediction method described in the first or second aspect.
[0043] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the channel success rate prediction method described in the first or second aspect.
[0044] The channel success rate prediction method, apparatus, electronic device, and computer program product provided in this application embodiment efficiently removes abnormal data by performing data cascade elimination on initial sample data, expands the sample based on the remaining first sample data, and provides effective artificial samples as second sample data; performs feature fusion and feature filtering on the first and second sample data to determine effective modeling features as target features; further, determines a better loss function as the target loss function based on the initial channel success rate obtained based on the target features, and constructs a channel success rate prediction model by combining the probability residual jump connection method. This can improve the prediction effect of the model and help improve the accuracy of the target channel success rate obtained by predicting the channel success rate based on the channel success rate prediction model and the channel data to be predicted. Attached Figure Description
[0045] To more clearly illustrate the technical solutions in this application 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 application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0046] Figure 1 This is a flowchart illustrating the channel success rate prediction method provided in the embodiments of this application;
[0047] Figure 2 This is a schematic diagram of a scenario for the probability residual jump connection strategy of the channel success rate prediction method provided in this application embodiment;
[0048] Figure 3 This is a schematic diagram of data cascading elimination in the channel success rate prediction method provided in the embodiments of this application;
[0049] Figure 4 This is a partial scenario diagram illustrating the sample expansion of the channel success rate prediction method provided in the embodiments of this application;
[0050] Figure 5 This is a schematic diagram illustrating the scenario of feature fusion and feature filtering in the channel success rate prediction method provided in this application embodiment;
[0051] Figure 6 This application presents a schematic diagram of the functional modules of an embodiment of the channel success rate prediction device.
[0052] Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation
[0053] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0054] Figure 1 A flowchart illustrating the channel success rate prediction method provided in this application embodiment. (Refer to...) Figure 1 This application provides a method for predicting channel success rate, which may include:
[0055] Step S100: Perform cascaded data removal based on the initial sample data to obtain the first sample data;
[0056] The channel success rate prediction method in this embodiment can be applied to electronic devices such as smartphones, tablets, PCs, and servers. In this embodiment, the electronic device can obtain channel prediction data features and their corresponding success rates as labeled data as initial sample data. The channels in this embodiment can be, but are not limited to, franchises, agents, e-commerce, etc., and the channel prediction data features can be as shown in Table 1 below:
[0057]
[0058]
[0059] Table 1
[0060] That is, the channel prediction data features may include, but are not limited to, string variables such as activity code, activity name, channel code, channel name, activity category code, activity category name, project category code, project category name, recommended product code, and recommended product name; int variables such as marketing cycle, number of recommended users, and total number of recommendations; and float variables such as recommendation rate, percentage of high-quality users, percentage of key personnel, percentage of users with more than 60 dissemination users, percentage of users who have conducted more than 3 business transactions in the past 3 months, percentage of users who have conducted business on this channel in the past 3 months, percentage of users who prefer this channel, percentage of young users, percentage of middle-aged users, and percentage of elderly users.
[0061] After obtaining the initial sample data, since this sample data often contains various invalid data (such as erroneous or meaningless data), this embodiment can perform cascaded data removal on the initial sample data to eliminate erroneous and invalid data. Specifically, this embodiment can first construct a joint distribution using each single feature variable and its corresponding success rate in the initial sample data as the target variable, and then use the joint distribution function to identify anomalous data from the initial sample data as initial anomalous data. It should be noted that the initial anomalous data only indicates that it is abnormal in certain dimensions. Therefore, this embodiment can construct a joint distribution of all feature variables and the target variable based on the initial anomalous data, and use the joint distribution function to identify data that is abnormal in the overall dimensions as target anomalous data. The target anomalous data is then removed from the initial sample data, and the remaining sample data is determined as the first sample data. This allows for subsequent sample expansion based on the first sample data to obtain the second sample data. By eliminating the influence of invalid data, the prediction accuracy can be improved.
[0062] Step S200: Based on the first sample data, perform sample augmentation to obtain the second sample data;
[0063] It should be noted that current GAN (Generative Adversarial Networks) data generation is based on a generator and a discriminator, but the generated data quality is not high. Therefore, this embodiment improves the original GAN by using a hybrid GAN data generation method. Specifically, this embodiment constructs multiple independently distributed generators, such as 3, 4, or 5, and in this embodiment, 3 are used. After constructing multiple generators, the first sample data is augmented using multiple generators and a discriminator. Specifically, the sample data generated by multiple generators can be predicted by the discriminator, and a loss value can be calculated based on the predicted value. The loss function of the original GAN is then corrected using the calculated loss value. The multiple generators are updated using the corrected loss function, and the updated generators generate more effective samples as second sample data based on the first sample data, thereby expanding the sample and improving the generalization ability of the subsequently constructed model.
[0064] It's important to note that a GAN framework has at least (but is not limited to) two components: a generative model (also called a generator) G and a discriminative model (also called a discriminator) D. During training, samples generated by the generative model and real samples are randomly sent to the discriminative model D in batches. The goal of the discriminative model D is to correctly identify real samples (output "true" or 1) and correctly identify generated samples, i.e., fake samples (output "false" or 0). The goal of the generative model is the opposite of the discriminative model: to minimize the probability that the discriminative model identifies it. Thus, G and D form a min-max game, where both continuously optimize themselves during training until equilibrium is reached—neither can improve further, meaning that fake samples and real samples are completely indistinguishable.
[0065] Step S300: Based on the first sample data and the second sample data, perform feature fusion and feature filtering to obtain the target features;
[0066] After obtaining the first and second sample data, due to feature redundancy in the original channel features, existing feature selection is based on the original features. This leads to the limitation that important features only become relevant after derivation. Therefore, this embodiment constructs a three-layer neural network, including a hidden layer. The hidden layer (hereinafter referred to as the hidden layer) is used to fuse the features of the first and second sample data, and further, the fused features are subjected to feature dimensionality reduction to achieve feature filtering, with the remaining features serving as target features. By fusing and filtering the features of the first and second sample data, effective variables can be obtained as target features.
[0067] Step S400: Based on the target features, predict the channel success rate to obtain an initial channel success rate, and determine the target loss function based on the initial channel success rate;
[0068] After obtaining the target features through feature fusion and feature selection, a model is built based on these target features. The channel success rate is then predicted using the constructed model in conjunction with the target features, and the predicted channel success rate is determined as the initial channel success rate. It is necessary to train the constructed model after its construction and before predicting the channel success rate. The specific training method can be the same as existing model training methods, and will not be elaborated upon in this embodiment.
[0069] Furthermore, the loss function of the constructed model is modified based on the initial channel success rate. Specifically, since the channel success rate is a probability value and exhibits polarization, and the mean squared error does not provide an ideal model fit when used as the loss function, this embodiment needs to redetermine the loss function for the channel success rate. The redetermined loss function in this embodiment needs to consider two aspects: first, the percentage of error; and second, ensuring consistency in magnitude between the predicted and actual values. Therefore, this embodiment can determine the percentage of error based on the initial channel success rate and the actual value corresponding to the target feature; and determine the ratio of the initial channel success rate to the actual value corresponding to the target feature; and determine the target loss function based on the percentage of error and the ratio. Modifying the loss function helps improve the prediction accuracy of the model based on the modified loss function.
[0070] Step S500: Construct a channel success rate prediction model based on the target loss function and the probability residual jump connection method, and predict the channel success rate based on the channel success rate prediction model and the channel data to be predicted to obtain the target channel success rate.
[0071] After obtaining the target loss function through loss function correction, this embodiment uses the aforementioned fusion features as modeling features and the target loss function as the loss function to design the framework of the channel success rate prediction model. Simultaneously, since the ResNet network employs skip connects (residual connections), this embodiment also uses a residual connection strategy when constructing the channel success rate prediction model. However, because the original residual connections connect regardless of whether the modules are valid, it easily leads to feature redundancy in the model. Furthermore, the skip connection intervals in the original residual connections are pre-set, which sacrifices the flexibility of skip connections. Therefore, after designing the model framework, this embodiment uses a probabilistic residual skip connection strategy for skip connections, thereby constructing the channel success rate prediction model.
[0072] Among them, the probabilistic residual jump connection strategy is as follows: Figure 2 As shown, Figure 2This is a schematic diagram illustrating the probabilistic residual jump connection strategy of the channel success rate prediction method provided in this application embodiment. Specifically, the constructed channel success rate prediction model is a neural network with three hidden layers. Starting from the fused features, jumps are made according to probabilities p11, p12, and p13, where p11 + p12 + p13 = 1; hidden layer 1 jumps according to probabilities p21 and p22, where p21 + p22 = 1; hidden layer 2 jumps according to p3 = 1, which is actually a guaranteed jump; hidden layer 3 is connected to the Loss (loss value), and there is no jump. It should be noted that the probability values in this embodiment are selected using a grid search strategy. By comparing the results of different groups of probability values, the optimal probability value can be obtained. Based on the optimal probability value, it can be determined whether the current channel success rate prediction model is the optimal model. If the current channel success rate prediction model is the optimal model, it is saved, and based on this optimal model combined with the channel data to be predicted, the channel success rate is predicted to obtain the target channel success rate. Specifically, the channel success rate prediction model is input into the optimal channel success rate prediction model. This model then predicts the channel success rate and sets the predicted value as the target channel success rate. By employing a probabilistic residual jump connection strategy, redundancy in jump connection features is resolved, achieving the effect of model integration and improving the model's generalization performance. This helps improve the accuracy of model-based predictions. The channel success rate prediction method provided in this application involves cascading data removal from the initial sample data to efficiently eliminate abnormal data. Based on the remaining first sample data, sample expansion is performed to provide effective artificial samples as second sample data. Feature fusion and feature filtering are performed on the first and second sample data to determine effective modeling features as target features. Furthermore, a better loss function is determined based on the initial channel success rate obtained from the target features as the target loss function. This, combined with the probabilistic residual jump connection method, constructs a channel success rate prediction model, which improves the model's prediction performance and enhances the accuracy of the target channel success rate predicted based on the channel success rate prediction model and the channel success rate prediction data.
[0073] In one embodiment, the step of performing cascaded data elimination based on initial sample data to obtain the first sample data includes:
[0074] Step S101: Construct a first joint distribution based on the single feature variable and the target variable in the initial sample data;
[0075] Step S102: Determine the target abnormal data based on the probability density of the first joint distribution;
[0076] Step S103: Remove the target abnormal data from the initial sample data to obtain the first sample data.
[0077] After obtaining the initial sample data, the prediction data features of each channel in the initial sample data are extracted as single feature variables, and the success rate in the initial sample data is used as the target variable. The single feature variables can be represented by x. i Let i = 1, 2, ..., m; the success rate can be represented by y; and a joint distribution is constructed based on the single feature variable and the target variable as the first joint distribution, specifically (x... i (y) represents a single feature variable x i The two-dimensional random variable and the success rate y are used as the first joint distribution.
[0078] After constructing the first joint distribution, its probability density is calculated. Abnormal data is initially identified from the initial sample data by comparing the probability density with a set probability threshold. This initially identified abnormal data is designated as initial abnormal data. It should be noted that the probability thresholds compared with the probability density of the first joint distribution and those subsequently compared with the probability density of the second joint distribution are both probability values set according to actual needs or environmental requirements. The two probability thresholds may be the same or different. For ease of distinction, the former can be referred to as the first probability threshold, and the latter as the second probability threshold. Since the initial abnormal data only indicates abnormality in certain dimensions, this embodiment also needs to construct a joint distribution of all feature variables and the target variable based on the initial abnormal data as a second joint distribution. The second joint distribution function is then used to further identify data with anomalies in the overall dimensions from the initial abnormal data as target abnormal data.
[0079] After identifying the target abnormal data, the data corresponding to the target abnormal data is found in the initial sample data and removed from the initial sample data. The remaining sample data is determined as the first sample data, so that the sample can be expanded based on the first sample data to obtain the second sample data. By removing the abnormal data in the initial sample data, the influence of invalid data can be eliminated and the prediction accuracy can be improved.
[0080] Further, the step of determining the target anomaly data based on the probability density of the first joint distribution includes:
[0081] Step S1021: Determine the initial abnormal data based on the probability density of the first joint distribution;
[0082] Step S1022: Construct a second joint distribution of the full feature variables and the target variable in the initial abnormal data;
[0083] Step S1023: Determine the target abnormal data based on the probability density of the second joint distribution.
[0084] After constructing the first joint distribution, calculate the probability density of the first joint distribution, specifically using p(x). i Let y) represent the probability density of the first joint distribution. Simultaneously, a pre-set and stored first probability threshold is obtained, and the probability density of the first joint distribution is compared with the first probability threshold. In this embodiment, the probability threshold can be α. i This means that if the probability density of the first joint distribution is less than the first probability density, then the data in the joint distribution are determined to be in x. i Anomalies were found in the y-dimension, and the location was in the x-dimension. i This involves combining all abnormal data along the y-axis to form an abnormal data set. In this embodiment, the abnormal data set can be represented by s. i This means that the union of the sets i = 1, 2, ..., m yields the abnormal data set s. u .
[0085] After identifying the initial outlier data, a second joint distribution is constructed based on the outlier data set, consisting of the joint distribution of all feature variables and the target variable. This second joint distribution can be represented as (x1, x2, x3, ..., x...). m Let , y) represent the total variable features, where the features are formed by multiple abnormal data points in the abnormal data set. The probability density of the second joint distribution is calculated. Based on the comparison between the probability density of the second joint distribution and the second probability threshold, data points exhibiting anomalies in the overall dimension are further identified from the initial abnormal data as target abnormal data. This allows for the identification of data corresponding to the target abnormal data from the initial sample data after identification, and their removal from the initial sample data. The remaining sample data is then designated as the first sample data. By removing abnormal data from the initial sample data, the influence of invalid data can be eliminated, improving prediction accuracy.
[0086] Further, the step of determining the target anomaly data based on the probability density of the second joint distribution includes:
[0087] Step S10231: Calculate the probability density of the second joint distribution;
[0088] Step S10232: Compare the probability density of the second joint distribution with a preset threshold;
[0089] Step S10233: If the probability density of the second joint distribution is less than the preset threshold, then the full feature variables in the second joint distribution are determined as target abnormal data.
[0090] After constructing the second joint distribution, the probability density of the second joint distribution is calculated, where the probability density of the second joint distribution can be represented by p(x1, x2, x3, ..., x...). mThe probability density of the second joint distribution is represented by y). Simultaneously, a preset second probability threshold (i.e., the preset threshold) is obtained, and the probability density of the second joint distribution is compared with the second probability threshold to determine the magnitude relationship between them. If, after comparison, it is determined that the probability density of the second joint distribution is less than the second probability threshold (i.e., the preset threshold), it indicates that the data in the second joint distribution is abnormal in the overall dimension. Therefore, the full feature variables in the second joint distribution that are abnormal in the overall dimension are identified as target abnormal data. In this embodiment, target abnormal data can be represented by sF. This facilitates finding data corresponding to the target abnormal data from the initial sample data after identifying the target abnormal data and removing it from the initial sample data. The remaining sample data is then identified as the first sample data. By removing abnormal data from the initial sample data, the influence of invalid data can be eliminated, improving prediction accuracy.
[0091] In one embodiment of this application, reference is made to Figure 3 , Figure 3 This diagram illustrates the cascaded elimination of data in the channel success rate prediction method provided in this application embodiment. Initial sample data is obtained, and the distribution is determined based on the first joint distribution (x1, y), (x2, y), ..., (x...). m The probability densities p(x1, y), p(x2, y), ..., p(x) corresponding to x1, y, ..., y) are p(x1, y), p(x2, y), ..., p(x3, y). m Whether the probability densities p(x1, y), p(x2, y), and p(xm) are less than the corresponding first probability thresholds α1, α2, and αm. m If y) is less than the corresponding first probability thresholds α1, α2, αm, then the corresponding x1, x2, xm will be set to... m The initial outlier data s1, s2, and sm are identified and formed into an outlier data set su. Based on this outlier data set, a second joint distribution (x1, x2, x3, ..., xm) is constructed. m , y), and calculate the probability density p(x1, x2, x3, ..., x) of the second joint distribution. m The probability density of the second joint distribution (s, y) is compared with the second probability threshold αu. If the probability density of the second joint distribution is less than the second probability threshold αu, then the full feature variables in the second joint distribution are identified as the target outlier data sF. This allows for the identification of the target outlier data, the search for corresponding data in the initial sample data, and the removal of this data from the initial sample data. The remaining sample data is then designated as the first sample data. By removing outlier data from the initial sample data, the influence of invalid data can be eliminated, improving prediction accuracy.
[0092] In one embodiment, the step of augmenting the sample data based on the first sample data to obtain the second sample data includes:
[0093] Step S201: Generate multiple intermediate sample data based on the first sample data using multiple initial generators;
[0094] Step S202: Summing the multiple intermediate sample data and inputting the summation into the discriminator to obtain the predicted value;
[0095] Step S203: Calculate the loss value based on the predicted value;
[0096] Step S204: Based on the loss value, correct the loss function of each initial generator to obtain multiple target generators;
[0097] Step S205: Based on the multiple target generators, the first sample data is augmented to obtain the second sample data.
[0098] After obtaining the first sample data, multiple generators are constructed as initial generators, each operating independently. In this embodiment, there can be three initial generators. Each of the three initial generators generates samples based on the first sample data, resulting in three sample data sets. These three sets are then assigned different weights w1, w2, and w3, which serve as intermediate sample data, where w1 + w2 + w3 = 1. After generating the three intermediate sample data sets, the sums are calculated, and the result is input into the discriminator for prediction, yielding a predicted value. Furthermore, the loss value is calculated by comparing the predicted value with the success rate corresponding to the original sample data.
[0099] Furthermore, the loss function corresponding to the initial generator in the original GAN is modified based on the calculated loss value. Specifically, the modified loss function is shown in the following formulas (1)-(3):
[0100] L1 = E X～P(X) [logD(x)] (1)
[0101]
[0102]
[0103] Where D represents the discriminator and G represents the generator. Formula (1) means that under the condition that X follows P(X), the result of the discriminator predicting X is calculated and the mean is taken; Formula (2) means that under the condition that Z follows P(Z), the result of the discriminator predicting and summing the results of the three generators is calculated and the mean of the logarithm of the difference between 1 and this summation result is taken; Formula (3) means that the overall loss function is the optimization of the maximum and minimum values of L1 and L2. The initial generator after the loss function is corrected is determined as the target generator, and the three target generators generate samples according to the first sample data respectively, generating three sample data as the second sample data. By generating more effective samples as the second sample data, the sample can be expanded, which is beneficial to improving the generalization ability of the model subsequently constructed.
[0104] In one example of this application, reference is made to Figure 4 , Figure 4 This diagram illustrates a partial scenario of sample expansion for the channel success rate prediction method provided in this application. Three initial generators, G1, G2, and G3, generate samples based on the first sample data z, resulting in three sample data sets. These sets are then assigned different weights w1, w2, and w3, which are used as intermediate sample data: w1*G1(z), w2*G2(z), and w3*G3(z), where w1+w2+w3=1. After generating these three intermediate sample data sets, they are summed, and the sum is input into the discriminator D for prediction, yielding a predicted value Pre. Furthermore, the loss value Loss is calculated by comparing the predicted value with the success rate corresponding to the original sample data.
[0105] In one embodiment, the step of performing feature fusion and feature filtering based on the first sample data and the second sample data to obtain the target features includes:
[0106] Step S301: Perform feature fusion on the first sample data and the second sample data to obtain fused features;
[0107] Step S302: Perform feature dimensionality reduction and filtering on the fused features based on preset regularization to obtain target features.
[0108] After obtaining the first and second sample data, a three-layer neural network is constructed, with one hidden layer. The dimensionality of the features in the hidden layer is consistent with the original feature dimension. Features from the first and second sample data can be fused through the hidden layer, specifically by concatenating the first and second sample data to obtain fused features. Further, a LASSO model is constructed based on the fused features, and a pre-defined L1 regularization is used to perform feature dimensionality reduction and selection. The remaining features after selection are determined as target features. The LASSO model, first proposed by Robert Tibshirani in 1996, stands for Least Absolute Shrinkage and Selection Operator, a regression model. This model is a compression estimation method that constructs a penalty function to obtain a more refined model, compressing some regression coefficients by forcing the sum of the absolute values of the coefficients to be less than a certain fixed value; simultaneously, some regression coefficients are set to zero, thus retaining the advantage of subset shrinkage. It is a biased estimator for handling data with multicollinearity. L1 regularization, also known as L1 weighting, refers to the sum of the absolute values of all elements in a weight vector w, typically represented as ||w||1. By fusing and filtering features from the first and second sample data, effective variables can be obtained as target features. This facilitates subsequent channel success rate prediction based on the target features, yielding an initial channel success rate, and determining the target loss function based on this initial success rate.
[0109] In one embodiment, the step of determining the target loss function based on the initial channel success rate includes:
[0110] Step S401: Determine the error percentage based on the initial channel success rate and the actual value corresponding to the target feature;
[0111] Step S402: Determine the ratio of the initial channel success rate to the actual value corresponding to the target feature;
[0112] Step S403: Determine the target loss function based on the error percentage and the ratio.
[0113] After obtaining the initial channel success rate, since the predicted value, i.e., the initial channel success rate, is a probabilistic problem, in order to reduce the impact of the polarization between the predicted and actual values on the error, this embodiment calculates the error percentage based on the actual value corresponding to the target feature from the initial channel success rate. The square root of both the predicted and actual values is taken to amplify their values, indirectly reducing the polarization. The absolute value of the difference between the square root of the actual and predicted values is then taken, and the ratio of this absolute value to the actual value is calculated to obtain the error percentage. Simultaneously, the ratio between the actual and predicted values is calculated to ensure consistency in their magnitudes, and the target loss function is formed by the product of the error percentage and this ratio.
[0114] The target loss function is shown in the following formula:
[0115]
[0116] in, This is a predicted value, specifically the initial channel success rate in this embodiment, y. i This is the actual value.
[0117] The target loss function, determined by the initial channel success rate, can improve the model's prediction accuracy.
[0118] In one example of this application, reference is made to Figure 5 , Figure 5 This is a schematic diagram illustrating the feature fusion and feature selection scenario of the channel success rate prediction method provided in this application embodiment. Specifically, in this embodiment, the first sample data and the second sample data, which serve as the original features, are fused through a hidden layer to obtain fused features as hidden layer features. The original features and the hidden layer features have the same dimension, which is m-dimensional in this embodiment. Further, it is determined whether the hidden layer features meet the dimensionality requirements. If they do, prediction is directly performed based on the hidden layer features to obtain the predicted value; if the hidden layer features do not meet the dimensionality requirements, feature dimensionality reduction and selection are performed on the hidden layer features using L1 regularization, and a model is built based on the selected features. Prediction is then performed based on the established model to obtain the predicted value (i.e., the initial channel success rate).
[0119] Furthermore, this application also provides a channel success rate prediction device.
[0120] Reference Figure 6 , Figure 6 This is a schematic diagram of the functional modules of an embodiment of the channel success rate prediction device of this application.
[0121] The channel success rate prediction device includes:
[0122] The elimination module 100 is used to perform cascaded elimination of data based on the initial sample data to obtain the first sample data;
[0123] The expansion module 200 is used to expand the sample data based on the first sample data to obtain the second sample data.
[0124] Feature processing module 300 is used to perform feature fusion and feature filtering based on the first sample data and the second sample data to obtain target features;
[0125] The determination module 400 is used to predict the channel success rate based on the target features, obtain an initial channel success rate, and determine a target loss function based on the initial channel success rate.
[0126] The construction module 500 is used to construct a channel success rate prediction model based on the target loss function and the probability residual jump connection method, so as to predict the channel success rate and obtain the target channel success rate based on the channel success rate prediction model and the channel data to be predicted.
[0127] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call a computer program in the memory 830 to execute the steps of the channel success rate prediction method, such as including:
[0128] Based on the initial sample data, cascaded data elimination is performed to obtain the first sample data;
[0129] Based on the first sample data, sample expansion is performed to obtain the second sample data;
[0130] Based on the first sample data and the second sample data, feature fusion and feature filtering are performed to obtain the target features;
[0131] Based on the target features, channel success rate is predicted to obtain an initial channel success rate, and a target loss function is determined based on the initial channel success rate.
[0132] A channel success rate prediction model is constructed based on the target loss function and the probability residual jump connection method. The channel success rate is then predicted based on the channel success rate prediction model and the channel data to be predicted, and the target channel success rate is obtained.
[0133] Furthermore, the logical instructions in the aforementioned memory 830 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 this application, in essence, 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 this application. 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.
[0134] On the other hand, embodiments of this application also provide 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 perform the steps of the channel success rate prediction method provided in the above embodiments, such as including:
[0135] Based on the initial sample data, cascaded data elimination is performed to obtain the first sample data;
[0136] Based on the first sample data, sample expansion is performed to obtain the second sample data;
[0137] Based on the first sample data and the second sample data, feature fusion and feature filtering are performed to obtain the target features;
[0138] Based on the target features, channel success rate is predicted to obtain an initial channel success rate, and a target loss function is determined based on the initial channel success rate.
[0139] A channel success rate prediction model is constructed based on the target loss function and the probability residual jump connection method. The channel success rate is then predicted based on the channel success rate prediction model and the channel data to be predicted, and the target channel success rate is obtained.
[0140] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program for causing a processor to perform the steps of the methods provided in the above embodiments, such as including:
[0141] Based on the initial sample data, cascaded data elimination is performed to obtain the first sample data;
[0142] Based on the first sample data, sample expansion is performed to obtain the second sample data;
[0143] Based on the first sample data and the second sample data, feature fusion and feature filtering are performed to obtain the target features;
[0144] Based on the target features, channel success rate is predicted to obtain an initial channel success rate, and a target loss function is determined based on the initial channel success rate.
[0145] A channel success rate prediction model is constructed based on the target loss function and the probability residual jump connection method. The channel success rate is then predicted based on the channel success rate prediction model and the channel data to be predicted, and the target channel success rate is obtained.
[0146] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).
[0147] 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.
[0148] 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.
[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application.
Claims
1. A method for predicting channel success rate, characterized in that, include: The first sample data is obtained by cascading data elimination based on the initial sample data. The initial sample data consists of channel prediction data features and their corresponding success rates as labeled data. Channels include franchising, agency, and e-commerce. The channel prediction data features include activity code, activity name, channel code, channel name, activity category code, activity category name, project category code, project category name, recommended product code, recommended product name, marketing cycle, number of recommended users, total number of recommendations, recommendation rate, percentage of high-quality users, percentage of key personnel, percentage of users with more than 60 dissemination users, percentage of users with more than 3 business transactions in the past 3 months, percentage of users who have conducted business on this channel in the past 3 months, percentage of users who prefer this channel, percentage of young users, percentage of middle-aged users, and percentage of elderly users. Based on the first sample data, sample expansion is performed to obtain the second sample data; Based on the first sample data and the second sample data, feature fusion and feature filtering are performed to obtain the target features; Based on the target features, channel success rate is predicted to obtain an initial channel success rate, and a target loss function is determined based on the initial channel success rate. A channel success rate prediction model is constructed based on the target loss function and the probability residual jump connection method. The channel success rate is then predicted based on the channel success rate prediction model and the channel data to be predicted, and the target channel success rate is obtained. The channel success rate prediction model is a neural network with three hidden layers. Starting from the fused features, it jumps according to probabilities p11, p12, and p13, where p11+p12+p13=1. Hidden layer 1 jumps according to probabilities p21 and p22, and p21+p22=1; hidden layer 2 jumps according to p3=1; hidden layer 3 of the neural network is connected to the target loss function.
2. The channel success rate prediction method according to claim 1, characterized in that, The step of performing cascaded data elimination based on the initial sample data to obtain the first sample data includes: Construct a first joint distribution based on the single feature variable and the target variable in the initial sample data; The target anomaly data is determined based on the probability density of the first joint distribution; The target abnormal data is removed from the initial sample data to obtain the first sample data.
3. The channel success rate prediction method according to claim 2, characterized in that, The step of determining the target anomaly data based on the probability density of the first joint distribution includes: The initial outlier data is determined based on the probability density of the first joint distribution; Construct a second joint distribution of all feature variables and the target variable in the initial abnormal data; The target anomaly data is determined based on the probability density of the second joint distribution.
4. The channel success rate prediction method according to claim 3, characterized in that, The step of determining the target anomaly data based on the probability density of the second joint distribution includes: Calculate the probability density of the second joint distribution; The probability density of the second joint distribution is compared with a preset threshold; If the probability density of the second joint distribution is less than the preset threshold, then all feature variables in the second joint distribution are identified as target abnormal data.
5. The channel success rate prediction method according to claim 1, characterized in that, The step of augmenting the sample data based on the first sample data to obtain the second sample data includes: Multiple intermediate sample data are generated from the first sample data using multiple initial generators; The summation of multiple intermediate sample data is then input into the discriminator to obtain the predicted value; Calculate the loss value based on the predicted value; Based on the loss value, the loss function of each initial generator is corrected to obtain multiple target generators; The first sample data is augmented using multiple target generators to obtain the second sample data.
6. The channel success rate prediction method according to claim 1, characterized in that, The step of performing feature fusion and feature filtering based on the first sample data and the second sample data to obtain the target features includes: The first sample data and the second sample data are fused to obtain fused features; The fused features are subjected to feature dimensionality reduction and filtering based on preset regularization to obtain the target features.
7. The channel success rate prediction method according to claim 1, characterized in that, The step of determining the target loss function based on the initial channel success rate includes: The error percentage is determined based on the initial channel success rate and the actual value corresponding to the target feature; Determine the ratio of the initial channel success rate to the actual value corresponding to the target feature; The target loss function is determined based on the error percentage and the ratio.
8. A channel success rate prediction device, characterized in that, include: The elimination module is used to perform cascaded elimination of data based on the initial sample data to obtain the first sample data. The initial sample data consists of channel prediction data features and their corresponding success rates as labeled data. Channels include franchising, agency, and e-commerce. The channel prediction data features include activity code, activity name, channel code, channel name, activity category code, activity category name, project category code, project category name, recommended product code, recommended product name, marketing cycle, number of recommended users, total number of recommendations, recommendation rate, percentage of high-quality users, percentage of key personnel, percentage of users with more than 60 dissemination users, percentage of users with more than 3 business transactions in the past 3 months, percentage of users who have conducted business on this channel in the past 3 months, percentage of users who prefer this channel, percentage of young users, percentage of middle-aged users, and percentage of elderly users. An expansion module is used to expand the sample data based on the first sample data to obtain the second sample data. The feature processing module is used to perform feature fusion and feature filtering based on the first sample data and the second sample data to obtain target features; The determination module is used to predict the channel success rate based on the target features, obtain an initial channel success rate, and determine the target loss function based on the initial channel success rate. The construction module is used to build a channel success rate prediction model based on the target loss function and the probability residual jump connection method, so as to predict the channel success rate and obtain the target channel success rate based on the channel success rate prediction model and the channel data to be predicted. The channel success rate prediction model is a neural network with three hidden layers. Starting from the fused features, it jumps according to probabilities p11, p12, and p13, where p11+p12+p13=1. Hidden layer 1 jumps according to probabilities p21 and p22, and p21+p22=1; hidden layer 2 jumps according to p3=1; hidden layer 3 of the neural network is connected to the target loss function; the target feature is obtained by feature selection based on the fused features.
9. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the channel success rate prediction method according to any one of claims 1 to 7.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the channel success rate prediction method according to any one of claims 1 to 7.