A marketing data prediction method based on multi-model fusion

Abstract

The application discloses a marketing data prediction method based on multi-model fusion, comprising the following steps: S1, a data preprocessing function is constructed, and an original feature set is formed through the data preprocessing function; S2, the original feature set in the step S1 is analyzed, the original feature set is divided into a training set and a test set, and standardization is performed; S3, three regression models are selected as basic models, and a stack model is constructed by taking a least absolute shrinkage and selection operator linear model as a meta model; S4, the stack model is trained by using the training set in the step S2, video watching transaction conversion rates and transaction amount prediction values in the test set are predicted, post-processing and evaluation are performed, and the processed stack model is saved; and S5, for a new data file, a prediction result is output based on the data preprocessing function constructed in the step S1 and the stack model of the prediction result trained in the step S4. The algorithm can effectively predict and analyze marketing related indexes.

Description

Technical Field

[0001] This invention belongs to the field of big data processing and model utilization, and is a marketing data prediction method based on multi-model fusion. Background Technology

[0002] In recent years, data sharing policies introduced in the industry have provided new infrastructure for predicting revenue from popular tourist attractions. By opening up core data interfaces such as real-time viewership and user profiles, it has become possible for predictive models to obtain front-end data input.

[0003] In terms of policy empowerment and data collection technology, current research mainly focuses on two types of policy tools: supply-side tools, which mandate that platforms provide at least 10 types of structured data fields; and environment-side tools (such as traffic subsidies and API call fee reductions), which reduce data acquisition costs. Zhang Rui (2023) used dynamic crawler technology to construct a policy-data response model, demonstrating that when the consistency of policy modeling improves by one unit, the data field completeness rate can be increased by 18.7%. Currently, cutting-edge data collection solutions have evolved to use a distributed real-time computing framework that integrates stream and batch processing, achieving second-level updates of metrics such as play counts and fan interactions through policy-authorized data pipelines.

[0004] In terms of predictive model building techniques, research focus has shifted from traditional econometric models to deep learning and multimodal fusion. Initially, autoregressive integral moving average models could only handle structured view count data. Wang et al. (2024) proposed a groundbreaking contrastive language-image pre-training-transformer neural network architecture, achieving cross-modal alignment between video content features and fan behavior sequences. Recent advancements involve using policy text as prior knowledge and adjusting model weights through attention mechanisms. For example, when the policy mentions the keyword "precision marketing," the model automatically enhances the weight coefficients of user geographic distribution characteristics.

[0005] The existing technology system has three key defects: First, the scope of data authorized by the policy has not yet covered new revenue channels such as live streaming rewards; second, cross-platform data synchronization is constrained by policy differences (e.g., the data openness of Douyin and Kuaishou differs by 37%); and third, policy lag causes the model cold start problem, with an average vacuum period of 11 days from policy promulgation to data interface readiness. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a marketing data prediction method based on multi-model fusion, which can combine existing data for analysis and predict marketing results.

[0007] The technical solution adopted in this invention is as follows: A marketing data prediction method based on multi-model fusion, comprising the following steps:

[0008] Step S1: Construct a data preprocessing function to read marketing-related data and transform the target variables to form the original feature set;

[0009] Step S2: Analyze the original feature set in Step S1, divide the original feature set into a training set and a test set, and standardize them;

[0010] Step S3: Select the extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model as the base models and adjust the parameters of the base models. Use the minimum absolute shrinkage and selection operator linear model as the meta-model to construct a stacked model.

[0011] Step S4: Train the stacked model using the training set from Step S2, predict the video viewing conversion rate and transaction amount prediction values ​​from the test set in Step S2, perform post-processing and evaluation, and save the processed stacked model.

[0012] Step S5: For the new data file, based on the data preprocessing function constructed in step S1 and the stacked model processed in step S4, the prediction result is output by standardizing the clustering features and the post-processing prediction results.

[0013] Furthermore, in step S1, the data preprocessing function reads marketing-related data and performs target variable transformation, specifically as follows:

[0014] Step S11: Read marketing-related data, including the number of effective video views, video sales performance level, number of Douyin account followers, user value clustering, video viewing conversion rate, and transaction amount;

[0015] Step S12: Use data preprocessing functions to obtain the logarithmic transformation value of transaction amount and fan play ratio characteristics;

[0016] Step S121: Calculate the logarithmic transformation value of the transaction amount, expressed by the formula:

[0017] ;

[0018] in, Indicates the original transaction amount; Represents the logarithmic form of the original transaction amount;

[0019] Step S122: Calculate the fan play ratio feature, as shown below:

[0020] ;

[0021] ;

[0022] in, This indicates the number of valid video playbacks after processing; eps represents the minimum positive number on the machine (eps≈2.22×). ); Indicates the number of valid video plays; Indicates the fan play ratio; This indicates the number of followers on a Douyin account.

[0023] Step S13: Extract the following data from the marketing-related data: number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, follower play ratio, video viewing conversion rate, and transaction amount, to form the original feature set.

[0024] Furthermore, in step S2, the original feature set from step S1 is divided into a training set and a test set and then standardized, specifically as follows:

[0025] Step S21: Divide the original feature set into a training set and a test set in an 8:2 ratio to obtain training set feature data, test set feature data, training set target data, and test set target data. The training set feature data consists of the number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, and follower-to-play ratio in the training set of the original feature set. The test set feature data consists of the number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, and follower-to-play ratio in the test set of the original feature set. The training set target data consists of the video viewing conversion rate and transaction amount in the training set of the original feature set. The test set target data consists of the video viewing conversion rate and transaction amount in the test set of the original feature set.

[0026] Step S22: Standardization of training set feature data: For the training set feature data, the transformation representation is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] in, denoted as the standardized feature value of the original data for a certain feature in the training set; X represents the original data for a certain feature in the training set; μ represents the mean of the original data for a certain feature in the training set; σ represents the standard deviation of the original data for a certain feature in the training set; n represents the number of samples of the original data for a certain feature in the training set. Represents the i-th sample of the original data for a certain feature in the training set; i is the index of the original data sample for a certain feature in the training set;

[0031] Step S23: Standardize the test set feature data: For the test set feature data, transform it into the following representation:

[0032] ;

[0033] ;

[0034] ;

[0035] in, The standardized feature values ​​of the original data for a certain feature in the test set; α represents the original data of a certain feature in the test set; α represents the mean of the original data of a certain feature in the test set on the training set; γ represents the standard deviation of the original data of a certain feature in the test set on the training set; m represents the number of samples of the original data of a certain feature in the test set. This represents the i-th sample of the original data for a certain feature in the test set;

[0036] Step S24: Save the standardization method from steps S22-S23 as a standardizer.

[0037] Furthermore, in step S3, the extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model are selected as the base models, and the base model parameters are set. A stacked model is constructed using the minimum absolute shrinkage and selection operator linear model as the meta-model; specifically:

[0038] Step S31: Construct the limiting gradient boosting regression model, with the parameters set as follows:

[0039] The Extreme Gradient Boosting Regression Model uses an Extreme Gradient Boosting Regressor, training a total of 300 trees; each tree has a depth of no more than 5 layers; the learning rate is 0.1; in each iteration, 80% of the sample data from the original feature set in the Extreme Gradient Boosting Regression Model is randomly used; the random seed is set to 42.

[0040] Step S32: Construct a random forest regression model with the following parameters:

[0041] The random forest regression model uses a random forest regressor to construct a total of 200 decision trees; the maximum depth of each tree is limited to 7 levels; each split node randomly selects 80% of the original feature set sample data to be used in the random forest regression model; the random seed is set to 42.

[0042] Step S33: Construct a lightweight gradient boosting machine regression model with the following parameters:

[0043] The lightweight gradient booster regression model uses a lightweight gradient booster regressor, training a total of 200 trees; the maximum depth of each tree is limited to 5 layers; the learning rate is set to 0.1; and the random seed is set to 42.

[0044] Step S34: Based on the minimum absolute shrinkage and the linear model of the selection operator as the meta-model, integrate the limit gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model with complete parameter settings from steps S1 to S3 to construct a stacked model.

[0045] Furthermore, in step S4, the stacked model is trained using the training set from step S2. The test set from step S2 predicts the video viewing conversion rate and transaction amount predictions in the test set, performs post-processing and evaluation, and saves the processed stacked model; specifically:

[0046] Step S41: Using the parameter-set extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model as the base models, and using the standardized training set feature data from step S22, perform training set target data prediction on the training set respectively to generate predicted values.

[0047] Step S42: Minimum absolute shrinkage and selection operator linear model as meta-model. Based on the predicted values ​​of the base model and the original feature set, output the final predicted training set target data, namely the video viewing conversion rate and transaction amount of the training set, and train the corresponding video viewing conversion rate stacked model and transaction amount stacked model.

[0048] Step S43: After training the stacked model, use the standardized test set feature data from step 23 to make predictions and obtain the predicted values ​​of the target data in the test set.

[0049] Step S44: Post-process the predicted video viewing conversion rate of the target data in the test set and evaluate the prediction effect;

[0050] Step S45: Post-process the predicted transaction amount of the target data in the test set and evaluate the prediction effect;

[0051] Step S46: Record the predicted range of the test set target data, namely the video viewing conversion rate and transaction amount of the test set, and save the processed stacked model.

[0052] Furthermore, in step S42, the formulas for predicting the target data of the training set, namely the video viewing conversion rate and transaction amount of the training set, are as follows:

[0053] Step S421: Predict the video view conversion rate, expressed by the formula:

[0054] ;

[0055] in, This represents the final predicted video view-to-conversion rate. The intercept of the meta-model is represented by ; j represents the index of the input feature for video viewing conversion rate; p represents the total number of predicted values ​​from the base model and the original features. The j-th video view conversion rate input feature is represented by the video view conversion rate prediction value of the three basic models and the five original feature values. The five original feature values ​​are: the number of effective video plays, the video sales performance level, the number of Douyin account followers, the user value clustering, and the follower play ratio. The weight coefficient represents the input feature of the conversion rate of the j-th video view. The weight coefficient is automatically learned by the minimum absolute contraction and selection operator linear model.

[0056] Step S422: Predict the transaction amount, expressed by the formula:

[0057] ;

[0058] ;

[0059] in, This represents the predicted transaction amount after logarithmic transformation; k is the index of the transaction amount input feature. This represents the k-th transaction amount input feature, which includes the transaction amount prediction values ​​of the 3 basic models and the original feature values; This represents the weight coefficient of the k-th transaction amount input feature; This represents the final predicted actual transaction amount.

[0060] Furthermore, in step S44, the predicted video viewing conversion rate of the target data in the test set is post-processed and the prediction effect is evaluated; specifically as follows:

[0061] Step S441: Post-processing based on the array cropping function to crop the predicted video viewing conversion rate to the [0,1] interval, represented as:

[0062] ;

[0063] in, This represents the predicted conversion rate of the video after cropping; test_pred represents the predicted conversion rate of the input video; a_min represents the minimum predicted conversion rate of the input video; a_max represents the maximum predicted conversion rate of the input video.

[0064] Step S442: Evaluate the prediction results by calculating the mean squared error function of the predicted conversion rate of the processed video viewing. The mean squared error function is calculated as follows:

[0065] ;

[0066] Where, mes represents the mean squared error of the test; r represents the number of videos in the test set; and t is the index used to traverse the videos in the test set. This represents the actual video view conversion rate of the t-th video; This represents the conversion rate of the t-th video after it has been predicted by the model and cropped.

[0067] Furthermore, in step S45, the predicted transaction amount of the target data in the test set is post-processed and the prediction effect is evaluated; specifically as follows:

[0068] Step S451: For the transaction amount target, perform an index conversion on the transaction amount prediction result to obtain the actual transaction amount prediction value;

[0069] Step S452: Use the array pruning function to ensure that the converted predicted value is non-negative, and obtain the post-processed final transaction amount prediction result;

[0070] Step S453: Calculate the mean square error of the processed transaction amount prediction result based on the mean square error function, and evaluate the performance of the stacked model using the value of the mean square error.

[0071] Furthermore, in step S5, for the new data file, based on the data preprocessing function constructed in step S1 and the stacked model processed in step S4, the prediction result is output through standardized clustering features and post-processing prediction results, specifically as follows:

[0072] Step S51: Load the normalizer saved in step S24 and the video viewing conversion rate stacked model and transaction amount stacked model trained in step S4, and create a copy of the input data for the new data file.

[0073] Step S52: Use the data preprocessing function in step S12 to obtain the logarithmic transformation value of the transaction amount and the fan playback ratio feature;

[0074] Step S53: Read the input data copy and extract clustering features: effective video play count, video sales performance level, number of Douyin account followers, and follower-to-play ratio. Use the normalizer saved in step S24 to standardize the clustering features and obtain standardized data.

[0075] Step S54: Input the standardized data from step S53 into the video viewing conversion rate stacking model and the transaction amount stacking model to obtain the prediction results;

[0076] Step S55: Based on steps S44 and S45, perform post-processing and evaluation of the video viewing conversion rate and transaction amount, and output the post-processed video viewing conversion rate and transaction amount.

[0077] Step S56: Enter the budget amount, and the estimated net profit will be automatically calculated, as shown below:

[0078] Projected net revenue = Video view conversion rate × Budgeted investment amount + Transaction amount.

[0079] The beneficial effects of this invention are as follows: The marketing data prediction method provided by this invention can deeply integrate existing marketing data such as the number of effective video views and the number of followers. By analyzing and extracting features such as video sales performance level and follower-view ratio, it can accurately predict marketing results such as conversion rate and sales amount. This invention optimizes and improves the prediction effect in many aspects: In data processing, standardized and unified feature dimensions are used; the model construction adopts the extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model as the basic models, and the minimum absolute contraction and selection operator linear model as the meta-model to integrate features; during training, a fixed random seed is used to ensure reproducibility; during prediction, an array pruning function is used to constrain the rationality of the results. These optimizations significantly improve the prediction accuracy and practicality, providing strong support for decisions such as marketing budget allocation and strategy adjustment. Attached Figure Description

[0080] Figure 1 This is a flowchart illustrating the steps of a marketing data prediction method based on multi-model fusion. Detailed Implementation

[0081] Reference Figure 1 A marketing data prediction method based on multi-model fusion includes the following steps:

[0082] Step S1: Construct a data preprocessing function to read marketing-related data and transform the target variables to form the original feature set;

[0083] Step S2: Analyze the original feature set in Step S1, divide the original feature set into a training set and a test set, and standardize them;

[0084] Step S3: Select the extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model as the base models and adjust the parameters of the base models. Use the minimum absolute shrinkage and selection operator linear model as the meta-model to construct a stacked model.

[0085] Step S4: Train the stacked model using the training set from Step S2, predict the video viewing conversion rate and transaction amount prediction values ​​from the test set in Step S2, perform post-processing and evaluation, and save the processed stacked model.

[0086] Step S5: For the new data file, based on the data preprocessing function constructed in step S1 and the stacked model processed in step S4, the prediction result is output by standardizing the clustering features and the post-processing prediction results.

[0087] Furthermore, in step S1, the data preprocessing function reads marketing-related data and performs target variable transformation, specifically as follows:

[0088] Step S11: Read marketing-related data, including the number of effective video views, video sales performance level, number of Douyin account followers, user value clustering, video viewing conversion rate, and transaction amount;

[0089] Step S12: Use data preprocessing functions to obtain the logarithmic transformation value of transaction amount and fan play ratio characteristics;

[0090] Step S121: Calculate the logarithmic transformation value of the transaction amount, expressed by the formula:

[0091] ;

[0092] in, Indicates the original transaction amount; Represents the logarithmic form of the original transaction amount;

[0093] Step S122: Calculate the fan play ratio feature, as shown below:

[0094] ;

[0095] ;

[0096] in, This indicates the number of valid video playbacks after processing; eps represents the minimum positive number on the machine (eps≈2.22×). ); Indicates the number of valid video plays; Indicates the fan play ratio; This indicates the number of followers on a Douyin account.

[0097] Step S13: Extract the following data from the marketing-related data: number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, follower play ratio, video viewing conversion rate, and transaction amount, to form the original feature set.

[0098] Furthermore, in step S2, the original feature set from step S1 is divided into a training set and a test set and then standardized, specifically as follows:

[0099] Step S21: Divide the original feature set into a training set and a test set in an 8:2 ratio to obtain training set feature data, test set feature data, training set target data, and test set target data. The training set feature data consists of the number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, and follower-to-play ratio in the training set of the original feature set. The test set feature data consists of the number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, and follower-to-play ratio in the test set of the original feature set. The training set target data consists of the video viewing conversion rate and transaction amount in the training set of the original feature set. The test set target data consists of the video viewing conversion rate and transaction amount in the test set of the original feature set.

[0100] Step S22: Standardization of training set feature data: For the training set feature data, the transformation representation is as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] in, denoted as the standardized feature value of the original data for a certain feature in the training set; X represents the original data for a certain feature in the training set; μ represents the mean of the original data for a certain feature in the training set; σ represents the standard deviation of the original data for a certain feature in the training set; n represents the number of samples of the original data for a certain feature in the training set. Represents the i-th sample of the original data for a certain feature in the training set; i is the index of the original data sample for a certain feature in the training set;

[0105] Step S23: Standardize the test set feature data: For the test set feature data, transform it into the following representation:

[0106] ;

[0107] ;

[0108] ;

[0109] in, The standardized feature values ​​of the original data for a certain feature in the test set; α represents the original data of a certain feature in the test set; α represents the mean of the original data of a certain feature in the test set on the training set; γ represents the standard deviation of the original data of a certain feature in the test set on the training set; m represents the number of samples of the original data of a certain feature in the test set. This represents the i-th sample of the original data for a certain feature in the test set;

[0110] Step S24: Save the standardization method from steps S22-S23 as a standardizer.

[0111] Furthermore, in step S3, the extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model are selected as the base models, and the base model parameters are set. A stacked model is constructed using the minimum absolute shrinkage and selection operator linear model as the meta-model; specifically:

[0112] Step S31: Construct the limiting gradient boosting regression model, with the parameters set as follows:

[0113] The Extreme Gradient Boosting Regression Model uses an Extreme Gradient Boosting Regressor, training a total of 300 trees; each tree has a depth of no more than 5 layers; the learning rate is 0.1; in each iteration, 80% of the sample data from the original feature set in the Extreme Gradient Boosting Regression Model is randomly used; the random seed is set to 42.

[0114] Step S32: Construct a random forest regression model with the following parameters:

[0115] The random forest regression model uses a random forest regressor to construct a total of 200 decision trees; the maximum depth of each tree is limited to 7 levels; each split node randomly selects 80% of the original feature set sample data to be used in the random forest regression model; the random seed is set to 42.

[0116] Step S33: Construct a lightweight gradient boosting machine regression model with the following parameters:

[0117] The lightweight gradient booster regression model uses a lightweight gradient booster regressor, training a total of 200 trees; the maximum depth of each tree is limited to 5 layers; the learning rate is set to 0.1; and the random seed is set to 42.

[0118] Step S34: Based on the minimum absolute shrinkage and the linear model of the selection operator as the meta-model, integrate the limit gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model with complete parameter settings from steps S1 to S3 to construct a stacked model.

[0119] Furthermore, in step S4, the stacked model is trained using the training set from step S2. The test set from step S2 predicts the video viewing conversion rate and transaction amount predictions in the test set, performs post-processing and evaluation, and saves the processed stacked model; specifically:

[0120] Step S41: Using the parameter-set extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model as the base models, and using the standardized training set feature data from step S22, perform training set target data prediction on the training set respectively to generate predicted values.

[0121] Step S42: Minimum absolute shrinkage and selection operator linear model as meta-model. Based on the predicted values ​​of the base model and the original feature set, output the final predicted training set target data, namely the video viewing conversion rate and transaction amount of the training set, and train the corresponding video viewing conversion rate stacked model and transaction amount stacked model.

[0122] Step S43: After training the stacked model, use the standardized test set feature data from step 23 to make predictions and obtain the predicted values ​​of the target data in the test set.

[0123] Step S44: Post-process the predicted video viewing conversion rate of the target data in the test set and evaluate the prediction effect;

[0124] Step S45: Post-process the predicted transaction amount of the target data in the test set and evaluate the prediction effect;

[0125] Step S46: Record the predicted range of the test set target data, namely the video viewing conversion rate and transaction amount of the test set, and save the processed stacked model.

[0126] Furthermore, in step S42, the formulas for predicting the target data of the training set, namely the video viewing conversion rate and transaction amount of the training set, are as follows:

[0127] Step S421: Predict the video view conversion rate, expressed by the formula:

[0128] ;

[0129] in, This represents the final predicted video view-to-conversion rate. The intercept of the meta-model is represented by ; j represents the index of the input feature for video viewing conversion rate; p represents the total number of predicted values ​​from the base model and the original features. The j-th video view conversion rate input feature is represented by the video view conversion rate prediction value of the three basic models and the five original feature values. The five original feature values ​​are: the number of effective video plays, the video sales performance level, the number of Douyin account followers, the user value clustering, and the follower play ratio. The weight coefficient represents the input feature of the conversion rate of the j-th video view. The weight coefficient is automatically learned by the minimum absolute contraction and selection operator linear model.

[0130] Step S422: Predict the transaction amount, expressed by the formula:

[0131] ;

[0132] ;

[0133] in, This represents the predicted transaction amount after logarithmic transformation; k is the index of the transaction amount input feature. This represents the k-th transaction amount input feature, which includes the transaction amount prediction values ​​of the 3 basic models and the original feature values; This represents the weight coefficient of the k-th transaction amount input feature; This represents the final predicted actual transaction amount.

[0134] Furthermore, in step S44, the predicted video viewing conversion rate of the target data in the test set is post-processed and the prediction effect is evaluated; specifically as follows:

[0135] Step S441: Post-processing based on the array cropping function to crop the predicted video viewing conversion rate to the [0,1] interval, represented as:

[0136] ;

[0137] in, This represents the predicted conversion rate of the video after cropping; test_pred represents the predicted conversion rate of the input video; a_min represents the minimum predicted conversion rate of the input video; a_max represents the maximum predicted conversion rate of the input video.

[0138] Step S442: Evaluate the prediction results by calculating the mean squared error function of the predicted conversion rate of the processed video viewing. The mean squared error function is calculated as follows:

[0139] ;

[0140] Where, mes represents the mean squared error of the test; r represents the number of videos in the test set; and t is the index used to traverse the videos in the test set. This represents the actual video view conversion rate of the t-th video; This represents the conversion rate of the t-th video after it has been predicted by the model and cropped.

[0141] Furthermore, in step S45, the predicted transaction amount of the target data in the test set is post-processed and the prediction effect is evaluated; specifically as follows:

[0142] Step S451: For the transaction amount target, perform an index conversion on the transaction amount prediction result to obtain the actual transaction amount prediction value;

[0143] Step S452: Use the array pruning function to ensure that the converted predicted value is non-negative, and obtain the post-processed final transaction amount prediction result;

[0144] Step S453: Calculate the mean square error of the processed transaction amount prediction result based on the mean square error function, and evaluate the performance of the stacked model using the value of the mean square error.

[0145] Furthermore, in step S5, for the new data file, based on the data preprocessing function constructed in step S1 and the stacked model processed in step S4, the prediction result is output through standardized clustering features and post-processing prediction results, specifically as follows:

[0146] Step S51: Load the normalizer saved in step S24 and the video viewing conversion rate stacked model and transaction amount stacked model trained in step S4, and create a copy of the input data for the new data file.

[0147] Step S52: Use the data preprocessing function in step S12 to obtain the logarithmic transformation value of the transaction amount and the fan playback ratio feature;

[0148] Step S53: Read the input data copy and extract clustering features: effective video play count, video sales performance level, number of Douyin account followers, and follower-to-play ratio. Use the normalizer saved in step S24 to standardize the clustering features and obtain standardized data.

[0149] Step S54: Input the standardized data from step S53 into the video viewing conversion rate stacking model and the transaction amount stacking model to obtain the prediction results;

[0150] Step S55: Based on steps S44 and S45, perform post-processing and evaluation of the video viewing conversion rate and transaction amount, and output the post-processed video viewing conversion rate and transaction amount.

[0151] Step S56: Enter the budget amount, and the estimated net profit will be automatically calculated, as shown below:

[0152] Projected net revenue = Video view conversion rate × Budgeted investment amount + Transaction amount.

[0153] A marketing data prediction method based on multi-model fusion includes the following steps:

[0154] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A marketing data prediction method based on multi-model fusion, characterized in that, Includes the following steps: Step S1: Construct a data preprocessing function to read marketing-related data and transform the target variables to form the original feature set; Step S2: Analyze the original feature set in Step S1, divide the original feature set into a training set and a test set, and standardize them; Step S3: Select the extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model as the base models and adjust the parameters of the base models. Use the minimum absolute shrinkage and selection operator linear model as the meta-model to construct a stacked model. Step S4: Train the stacked model using the training set from Step S2. Predict the video viewing conversion rate and transaction amount from the test set in Step S2, perform post-processing and evaluation, and save the processed stacked model; specifically: Step S41: Using the parameter-set extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model as the base models, and using the standardized training set feature data, perform training set target data prediction on the training set respectively to generate predicted values; Step S42: Minimum absolute shrinkage and selection operator linear model as meta-model. Based on the predicted values ​​of the base model and the original feature set, output the final predicted training set target data, namely the video viewing conversion rate and transaction amount of the training set, and train the corresponding video viewing conversion rate stacked model and transaction amount stacked model. Step S43: After training the stacked model, use the standardized test set feature data to make predictions and obtain the predicted values ​​of the target data in the test set. Step S44: Post-process the predicted video view conversion rate of the target data in the test set and evaluate the prediction effect; the details are as follows: Step S441: Post-processing based on the array cropping function to crop the predicted video viewing conversion rate to the [0,1] interval, represented as: ； in, This represents the predicted conversion rate of the video after cropping; test_pred represents the predicted conversion rate of the input video; a_min represents the minimum predicted conversion rate of the input video; a_max represents the maximum predicted conversion rate of the input video. Step S442: Evaluate the prediction results by calculating the mean square error function of the predicted conversion rate of the processed video viewing. Step S45: Post-process the predicted transaction amount of the target data in the test set and evaluate the prediction effect; Step S46: Record the predicted range of the test set target data, namely the video viewing conversion rate and transaction amount of the test set, and save the processed stacked model; Step S5: For the new data file, based on the data preprocessing function constructed in step S1 and the stacked model processed in step S4, the prediction result is output by standardizing the clustering features and the post-processing prediction results.

2. The marketing data prediction method based on multi-model fusion according to claim 1, characterized in that, In step S1, the data preprocessing function reads marketing-related data and performs target variable transformation, specifically as follows: Step S11: Read marketing-related data, including the number of effective video views, video sales performance level, number of Douyin account followers, user value clustering, video viewing conversion rate, and transaction amount; Step S12: Use data preprocessing functions to obtain the logarithmic transformation value of transaction amount and fan play ratio characteristics; Step S121: Calculate the logarithmic transformation value of the transaction amount, expressed by the formula: ； in, Indicates the original transaction amount; Represents the logarithmic form of the original transaction amount; Step S122: Calculate the fan play ratio feature, as shown below: ； ； in, This indicates the number of valid playbacks of the processed video; eps represents the minimum positive number on the machine. Indicates the number of valid video plays; Indicates the fan-to-view ratio; This indicates the number of followers on a Douyin account. Step S13: Extract the following data from the marketing-related data: number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, follower play ratio, video viewing conversion rate, and transaction amount, to form the original feature set.

3. The marketing data prediction method based on multi-model fusion according to claim 2, characterized in that, In step S2, the original feature set from step S1 is divided into a training set and a test set and then standardized, specifically as follows: Step S21: Divide the original feature set into a training set and a test set in an 8:2 ratio to obtain training set feature data, test set feature data, training set target data, and test set target data. The training set feature data consists of the number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, and follower-to-play ratio in the training set of the original feature set. The test set feature data consists of the number of effective video plays, video sales performance level, number of Douyin account followers, user value clustering, and follower-to-play ratio in the test set of the original feature set. The training set target data consists of the video viewing conversion rate and transaction amount in the training set of the original feature set. The test set target data consists of the video viewing conversion rate and transaction amount in the test set of the original feature set. Step S22: Standardization of training set feature data: For the training set feature data, the transformation representation is as follows: ； ； ； in, denoted as the standardized feature value of the original data for a certain feature in the training set; X represents the original data for a certain feature in the training set; μ represents the mean of the original data for a certain feature in the training set; σ represents the standard deviation of the original data for a certain feature in the training set; n represents the number of samples of the original data for a certain feature in the training set. Represents the i-th sample of the original data for a certain feature in the training set; i is the index of the original data sample for a certain feature in the training set; Step S23: Standardize the test set feature data: For the test set feature data, transform it into the following representation: ； ； ； in, This represents the standardized feature value of the original data for a certain feature in the test set. α represents the original data of a certain feature in the test set; α represents the mean of the original data of a certain feature in the test set on the training set; γ represents the standard deviation of the original data of a certain feature in the test set on the training set; m represents the number of samples of the original data of a certain feature in the test set. This represents the i-th sample of the original data for a certain feature in the test set; Step S24: Save the standardization method from steps S22-S23 as a standardizer.

4. The marketing data prediction method based on multi-model fusion according to claim 3, characterized in that, In step S3, the extreme gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model are selected as the basic models, and the basic model parameters are set. A stacked model is constructed using the minimum absolute shrinkage and selection operator linear model as the meta-model; specifically: Step S31: Construct the limiting gradient boosting regression model, with the parameters set as follows: The Extreme Gradient Boosting Regression Model uses an Extreme Gradient Boosting Regressor, training a total of 300 trees; each tree has a depth of no more than 5 layers; the learning rate is 0.1; in each iteration, 80% of the sample data from the original feature set in the Extreme Gradient Boosting Regression Model is randomly used; the random seed is set to 42. Step S32: Construct a random forest regression model with the following parameters: The random forest regression model uses a random forest regressor to construct a total of 200 decision trees; the maximum depth of each tree is limited to 7 levels; each split node randomly selects 80% of the original feature set sample data to be used in the random forest regression model; the random seed is set to 42. Step S33: Construct a lightweight gradient boosting machine regression model with the following parameters: The lightweight gradient booster regression model uses a lightweight gradient booster regressor, training a total of 200 trees; the maximum depth of each tree is limited to 5 layers; the learning rate is set to 0.1; and the random seed is set to 42. Step S34: Based on the minimum absolute shrinkage and selection operator linear model as the meta-model, integrate the limit gradient boosting regression model, random forest regression model, and lightweight gradient boosting machine regression model with complete parameter settings from steps S31 to S33 to construct a stacked model.

5. The marketing data prediction method based on multi-model fusion according to claim 4, characterized in that: In step S42, the formulas for predicting the target data of the training set, namely the video viewing conversion rate and transaction amount of the training set, are as follows: Step S421: Predict the video view conversion rate, expressed by the formula: ； in, This represents the final predicted video view-to-conversion rate. The intercept of the meta-model is represented by ; j represents the index of the input feature for video viewing conversion rate; p represents the total number of predicted values ​​from the base model and the original features. The j-th video view conversion rate input feature is represented by the video view conversion rate prediction value of the three basic models and the five original feature values. The five original feature values ​​are: the number of effective video plays, the video sales performance level, the number of Douyin account followers, the user value clustering, and the follower play ratio. The weight coefficient represents the input feature of the conversion rate of the j-th video view. The weight coefficient is automatically learned by the minimum absolute contraction and selection operator linear model. Step S422: Predict the transaction amount, expressed by the formula: ； ； in, This represents the predicted transaction amount after logarithmic transformation; k is the index of the transaction amount input feature. This represents the k-th transaction amount input feature, which includes the transaction amount prediction values ​​of the 3 basic models and the original feature values; This represents the weight coefficient of the k-th transaction amount input feature; This represents the final predicted actual transaction amount.

6. The marketing data prediction method based on multi-model fusion according to claim 5, characterized in that: Step S442: Evaluate the prediction results by calculating the mean squared error function of the predicted conversion rate of the processed video viewing. The mean squared error function is calculated as follows: ； Where, mes represents the mean squared error of the test; r represents the number of videos in the test set; and t is the index used to traverse the videos in the test set. This represents the actual video view conversion rate of the t-th video; This represents the conversion rate of the t-th video after it has been predicted by the model and cropped.

7. The marketing data prediction method based on multi-model fusion according to claim 6, characterized in that: In step S45, the predicted transaction amount of the target data in the test set is post-processed and the prediction effect is evaluated; specifically as follows: Step S451: For the transaction amount target, perform an index conversion on the transaction amount prediction result to obtain the actual transaction amount prediction value; Step S452: Use the array pruning function to ensure that the converted predicted value is non-negative, and obtain the post-processed final transaction amount prediction result; Step S453: Calculate the mean square error of the processed transaction amount prediction result based on the mean square error function, and evaluate the performance of the stacked model using the value of the mean square error.

8. The marketing data prediction method based on multi-model fusion according to claim 7, characterized in that: In step S5, for the new data file, based on the data preprocessing function constructed in step S1 and the stacked model processed in step S4, the prediction result is output through standardized clustering features and post-processing prediction results, specifically: Step S51: Load the normalizer saved in step S24 and the video viewing conversion rate stacked model and transaction amount stacked model trained in step S4, and create a copy of the input data for the new data file. Step S52: Use the data preprocessing function in step S12 to obtain the logarithmic transformation value of the transaction amount and the fan playback ratio feature; Step S53: Read the input data copy and extract clustering features: effective video play count, video sales performance level, number of Douyin account followers, and follower-to-play ratio. Use the normalizer saved in step S24 to standardize the clustering features and obtain standardized data. Step S54: Input the standardized data from step S53 into the video viewing conversion rate stacking model and the transaction amount stacking model to obtain the prediction results; Step S55: Based on steps S44 and S45, perform post-processing and evaluation of the video viewing conversion rate and transaction amount, and output the post-processed video viewing conversion rate and transaction amount. Step S56: Enter the budget amount, and the estimated net profit will be automatically calculated, as shown below: Projected net revenue = Video view conversion rate × Budgeted investment amount + Transaction amount.

Images