CN-121981758-A - Conversion rate acquisition method, conversion rate acquisition device, conversion rate acquisition equipment and storage medium

Abstract

The embodiment of the specification provides a conversion rate acquisition method, a device, equipment and a storage medium, wherein the method comprises the steps of determining a crowd to which a user belongs according to user information, acquiring time sequence data of preset behaviors executed by the crowd in a first period taking a current date as an end date, searching historical time sequence data matched with the time sequence data in the first period in a crowd historical behavior time sequence library corresponding to the crowd, determining the end date of the searched historical time sequence data, determining conversion rate reference time sequence data according to the conversion rate of the crowd in a period after the end date, and inputting the conversion rate reference time sequence data and the user information into a conversion rate prediction model trained in advance so as to acquire the conversion rate predicted by the conversion rate prediction model.

Inventors

• ZHOU YU
• XIE JINCHENG
• AI XIAOYAN
• WANG WEI

Assignees

• 重庆蚂蚁消费金融有限公司

Dates

Publication Date

20260505

Application Date

20260327

Claims (20)

1. 1. A conversion acquisition method, the method comprising: Determining a target crowd of a target user based on target user information of the target user, and acquiring time sequence data of the target crowd for executing preset behaviors in a first period by taking a current date as a first end date of the first period; Determining historical time sequence data corresponding to a plurality of second periods in a crowd historical behavior time sequence library corresponding to the target crowd, determining target historical time sequence data matched with the time sequence data from the plurality of historical time sequence data, wherein the period duration of the second periods is the same as the period duration of the first periods before the first periods; determining a second end date of the target historical time sequence data, taking the date after the second end date as a first start date of a third period, and acquiring conversion rate reference time sequence data in the third period; And inputting the conversion rate reference time sequence data and the target user information into a conversion rate prediction model, and acquiring the conversion rate of the target user in a fourth period based on the conversion rate prediction model, wherein the second starting date of the fourth period is the later date of the current date, and the period duration of the fourth period is the same as the period duration of the third period.
2. 2. The method according to claim 1, wherein the determining the target crowd to which the target user belongs based on the target user information of the target user, taking the current date as the first end date of the first period, and obtaining the time sequence data of the target crowd for executing the preset behavior in the first period, includes: acquiring target user information of a target user; determining a target crowd to which the target user belongs based on the target user information; Determining a current date as a first end date, determining a first history date of a first period duration before the current date as a third start date, and determining a first period based on the third start date and the first end date; And acquiring time sequence data of the target crowd for executing preset behaviors in the first period.
3. 3. The method of claim 1, wherein before determining historical time series data corresponding to the second plurality of periods from the crowd-history behavior time series library corresponding to the target crowd, determining target historical time series data matching the time series data from the plurality of historical time series data, further comprises: dividing users into a plurality of categories of people based on user information; Acquiring the ratio of the number of people executing the preset behaviors in the target crowd to the total number of people in the target crowd on the same day on each second historical date before the current date, and determining the ratio as the behavior probability of the target crowd executing the preset behaviors on the same day; And constructing a crowd historical behavior time sequence base for executing the preset behaviors by the target crowd based on the date time sequence of each second historical date and the behavior probability of executing the preset behaviors in each second historical date.
4. 4. The method of claim 1, wherein determining historical time series data corresponding to a plurality of second periods in the crowd-historic behavior time series library corresponding to the target crowd, and determining target historical time series data matching the time series data from the plurality of historical time series data, comprises: Determining each third historical date before the first period as a third final date, and determining a fourth historical date of the first period duration before each third final date as a fourth starting date; Determining a plurality of second periods based on each of the fourth endpoint dates and the corresponding third endpoint dates; acquiring historical time sequence data corresponding to each second period from a crowd historical behavior time sequence library corresponding to the target crowd; determining the similarity of the time sequence data and each historical time sequence data; And determining the maximum similarity of the preset number in the similarities as target similarity, and determining the historical time sequence data corresponding to the target similarity as target historical time sequence data matched with the time sequence data.
5. 5. The method of claim 4, wherein determining a second end date of the target historical time series data, taking a date subsequent to the second end date as a first start date of a third period, obtaining conversion rate reference time series data in the third period, comprises: Determining a second end date of each of the target historical time series data; Determining a later date of each second endpoint date as a first endpoint date, and determining a fifth historical date of a second period duration after each first endpoint date as a fourth endpoint date; determining a plurality of third periods based on each of the first endpoint dates and the corresponding fourth endpoint dates; Acquiring conversion rate time sequence data of the target crowd in each third period; and carrying out averaging treatment on the plurality of conversion rate time sequence data to obtain conversion rate reference time sequence data in the third period.
6. 6. The method of claim 1, wherein the inputting the conversion reference time series data and the target user information into a conversion prediction model, and obtaining the conversion of the target user in a fourth period based on the conversion prediction model, comprises: Inputting the conversion rate reference time sequence data and the target user information into a conversion rate prediction model to obtain a first hidden vector corresponding to the conversion rate reference time sequence data and a second hidden vector corresponding to the target user information; performing splicing treatment on the first hidden vector and the second hidden vector to obtain a spliced vector; and generating the conversion rate of the target user in a fourth period based on the splicing vector by adopting the conversion rate prediction model.
7. 7. The method of claim 6, wherein the conversion rate prediction model includes a fully connected neural network and a feature cross network, wherein the inputting the conversion rate reference time series data and the target user information into the conversion rate prediction model to obtain a first hidden vector corresponding to the conversion rate reference time series data, and wherein the second hidden vector corresponding to the target user information includes: Inputting the conversion rate reference time sequence data and the target user information into a conversion rate prediction model; and acquiring a first hidden vector corresponding to the conversion rate reference time sequence data based on the fully connected neural network, and acquiring a second hidden vector corresponding to the target user information based on the characteristic crossover network.
8. 8. A method of training a conversion rate prediction model, the method comprising: acquiring sample user information of a sample user, acquiring a conversion state label of the sample user in a first sample period, and determining a sample target crowd to which the sample user belongs based on the sample user information; Taking a sample date as a first sample end date of a second sample period, acquiring sample time sequence data of the sample target crowd for executing preset actions in the second sample period, wherein the sample date is a previous date of the first sample end date of the first sample period; Determining historical sample time sequence data corresponding to a plurality of third sample periods in a crowd historical behavior time sequence library corresponding to the sample target crowd, determining target historical sample time sequence data matched with the sample time sequence data from the plurality of historical sample time sequence data, wherein the period duration of the third sample periods is the same as the period duration of the second sample periods before the second sample periods; determining a second sample end date of the target historical sample time sequence data, taking the date after the second sample end date as a second sample start date of a fourth sample period, and acquiring sample conversion rate reference time sequence data in the fourth sample period; Constructing a training sample set based on the conversion state label, the sample conversion rate reference time sequence data and the sample user information; Inputting a target training sample into a conversion rate prediction model to be trained, so that the conversion rate prediction model outputs the predicted conversion rate of the sample user in the first sample period based on the sample conversion rate reference time sequence data and the sample user information, wherein the target training sample is any training sample in the training sample set; determining a loss function value based on the predicted conversion and the conversion status label; Model parameters of the conversion rate prediction model are updated based on the loss function values to train the conversion rate prediction model.
9. 9. The method of claim 8, wherein the obtaining the sample user's conversion status tag over a first sample period comprises: determining a later sample date of a sample date as a first sample start date, and determining a first sample history date of a second period duration after the first sample start date as a third sample end date; determining a first sample period based on the first sample start date and the third sample end date; And acquiring a conversion state label of the sample user in the first sample period.
10. 10. The method of claim 8, wherein the obtaining sample timing data for the sample target population to perform a predetermined action during a second sample period with the sample date being a first sample endpoint date of the second sample period comprises: determining a sample date as a first sample end date, and determining a second sample history date of a first period duration before the sample date as a third sample start date; Determining a second sample period based on the third sample start date and the first sample end date; And acquiring sample time sequence data of the sample target crowd for executing preset behaviors in the second sample period.
11. 11. The method of claim 8, wherein determining historical sample timing data corresponding to a plurality of third sample periods in the population historical behavior timing library corresponding to the sample target population, and determining target historical sample timing data matching the sample timing data from the plurality of historical sample timing data, comprises: determining each third sample history date before the second sample period as a fourth sample end date, and determining each fourth sample history date of the first period duration before the fourth sample end date as a fourth sample start date; Determining a plurality of third sample periods based on each of the fourth sample start dates and the corresponding fourth sample end dates; Acquiring historical sample time sequence data corresponding to each third sample period from a crowd historical behavior time sequence library corresponding to the sample target crowd; Determining sample similarity of the sample time sequence data and each historical sample time sequence data; And determining the maximum sample similarity of the preset number in the sample similarity as target sample similarity, and determining the historical sample time sequence data corresponding to the target sample similarity as target historical sample time sequence data matched with the sample time sequence data.
12. 12. The method of claim 8, wherein the determining a second sample end date of the target historical sample timing data, taking a date subsequent to the second sample end date as a second sample start date of a fourth sample period, obtaining sample conversion reference timing data within the fourth sample period, comprises: determining a second sample end date of each of the target historical sample timing data; Determining a later sample date of each second sample end date as a fifth sample start date, and determining a fifth sample history date of a second period duration after each fifth sample start date as a fifth sample end date; Determining a plurality of fourth sample periods based on each of the fifth sample start dates and the corresponding fifth sample end dates; acquiring sample conversion rate time sequence data of the sample target crowd in each fourth sample period; And carrying out averaging treatment on the plurality of sample conversion rate time sequence data to obtain sample conversion rate reference time sequence data in the fourth sample period.
13. 13. The method of claim 8, wherein inputting the target training sample into a conversion prediction model to be trained such that the conversion prediction model outputs a predicted conversion of the sample user during the first sample period based on the sample conversion reference time series data and the sample user information comprises: Inputting a target training sample into a conversion rate prediction model to be trained; acquiring a first sample hidden vector corresponding to the sample conversion reference time sequence data based on the conversion rate prediction model, and acquiring a second sample hidden vector corresponding to the sample user information; performing splicing treatment on the first sample hidden vector and the second sample hidden vector to obtain a sample spliced vector; and predicting the predicted conversion rate of the sample user in the first sample period based on the sample splicing vector.
14. 14. The method of claim 13, wherein the conversion rate prediction model includes a fully connected neural network and a feature cross network, wherein the obtaining a first sample hidden vector corresponding to the sample conversion rate reference time series data based on the conversion rate prediction model, and obtaining a second sample hidden vector corresponding to the sample user information, comprises: and acquiring a first sample hidden vector corresponding to the sample conversion rate reference time sequence data based on the fully connected neural network, and acquiring a second sample hidden vector corresponding to the sample user information based on the characteristic crossover network.
15. 15. The method of claim 8, wherein the determining a loss function value based on the predicted conversion and the conversion status tag comprises: a loss function value between the predicted conversion and the conversion state label is determined based on a cross entropy loss function.
16. 16. The method of claim 8, wherein the updating model parameters of the conversion rate prediction model based on the loss function value to train the conversion rate prediction model comprises: Updating model parameters of the conversion rate prediction model based on the loss function value, and judging whether the updated conversion rate prediction model meets a preset convergence condition or not; If the conversion rate prediction model meets the preset convergence condition, a trained conversion rate prediction model is obtained; And if the conversion rate prediction model does not meet the preset convergence condition, transferring to the step of inputting the target training sample into the conversion rate prediction model to be trained until the conversion rate prediction model meets the preset convergence condition, and obtaining a trained conversion rate prediction model.
17. 17. A conversion rate obtaining apparatus, characterized in that the apparatus comprises: The behavior data acquisition unit is used for determining a target crowd to which the target user belongs based on target user information of the target user, and acquiring time sequence data of the target crowd for executing preset behaviors in a first period by taking the current date as a first end date of the first period; A behavior data matching unit, configured to determine historical time sequence data corresponding to a plurality of second periods in a crowd historical behavior time sequence library corresponding to the target crowd, determine target historical time sequence data matched with the time sequence data from the plurality of historical time sequence data, where a period duration of the second period is the same as a period duration of the first period before the first period; A reference data obtaining unit, configured to determine a second end date of the target historical time series data, and obtain conversion rate reference time series data in a third period by taking a date subsequent to the second end date as a first start date of the third period; The conversion rate prediction unit is configured to input the conversion rate reference time sequence data and the target user information into a conversion rate prediction model, obtain a conversion rate of the target user in a fourth period based on the conversion rate prediction model, where a second starting date of the fourth period is a date subsequent to the current date, and a period duration of the fourth period is the same as a period duration of the third period.
18. 18. A training device for a conversion rate prediction model, the device comprising: The first acquisition unit is used for acquiring sample user information of a sample user, acquiring a conversion state label of the sample user in a first sample period, and determining a sample target crowd to which the sample user belongs based on the sample user information; The second acquisition unit is used for acquiring sample time sequence data of the sample target crowd for executing preset actions in a second sample period by taking the sample date as a first sample end date of the second sample period, wherein the sample date is a date before the first sample start date of the first sample period; A sample data matching unit, configured to determine historical sample time sequence data corresponding to a plurality of third sample periods in a crowd historical behavior time sequence library corresponding to the sample target crowd, determine target historical sample time sequence data matched with the sample time sequence data from the plurality of historical sample time sequence data, where a period duration of the third sample period is the same as a period duration of the second sample period before the second sample period; a third obtaining unit, configured to determine a second sample end date of the target historical sample timing data, and obtain sample conversion rate reference timing data in a fourth sample period by using a date subsequent to the second sample end date as a second sample start date of the fourth sample period; the training set construction unit is used for constructing a training sample set based on the conversion state label, the sample conversion rate reference time sequence data and the sample user information; The model training unit is used for inputting a target training sample into a conversion rate prediction model to be trained, so that the conversion rate prediction model outputs the predicted conversion rate of the sample user in the first sample period based on the sample conversion rate reference time sequence data and the sample user information, and the target training sample is any training sample in the training sample set; a loss value calculation unit for determining a loss function value based on the predicted conversion rate and the conversion state label; And a parameter updating unit for updating model parameters of the conversion rate prediction model based on the loss function value to train the conversion rate prediction model.
19. 19. A computer device comprising a processor and a memory, wherein the memory stores a computer program adapted to be loaded by the processor and to perform the steps of the method according to any of claims 1 to 16.
20. 20. A storage medium storing a computer program which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 16.

Description

Conversion rate acquisition method, conversion rate acquisition device, conversion rate acquisition equipment and storage medium Technical Field The present description relates to the field of computer technology, and more particularly, to a conversion rate acquisition method, apparatus, device, and storage medium in the field of computer technology. Background Conversion rate is an important index in the fields of computer science, data science and marketing, and the core goal is to predict the probability that a user will complete a certain target behavior (such as purchasing goods, clicking advertisements, registering account numbers, etc.) in a specific scene. Along with the rapid development of big data and artificial intelligence technology, the conversion rate prediction gradually evolves from a traditional statistical learning method to a deep learning method driven by data and a model, so that the accuracy of the conversion rate prediction is improved to a certain extent. However, in some transaction scenarios, the transformation behavior of the user presents dual features of periodicity and sparsity, and the deep learning model in the related art has few learnable positive samples, so that the sparse periodic rule in the scenario cannot be fully modeled, and the generalization capability of the model is weakened, and the prediction accuracy in the scenario is reduced. Disclosure of Invention The specification provides a conversion rate acquisition method, a conversion rate acquisition device, conversion rate acquisition equipment and a storage medium, wherein the conversion rate acquisition method can improve accuracy of conversion rate prediction and generalization capability of a conversion rate prediction model. In a first aspect, embodiments of the present disclosure provide a conversion obtaining method, the method including: Determining a target crowd to which the target user belongs based on target user information of the target user, and acquiring time sequence data of the target crowd for executing preset behaviors in a first period by taking a current date as a first end date of the first period; determining historical time sequence data corresponding to a plurality of second periods in a crowd historical behavior time sequence library corresponding to the target crowd, and determining target historical time sequence data matched with the time sequence data from the plurality of historical time sequence data, wherein the period duration of the second periods is the same as the period duration of the first periods before the first periods; Determining a second end date of the target historical time sequence data, taking the date after the second end date as a first start date of a third period, and acquiring conversion rate reference time sequence data in the third period; and inputting the conversion rate reference time sequence data and the target user information into a conversion rate prediction model, and acquiring the conversion rate of the target user in a fourth period based on the conversion rate prediction model, wherein the second starting date of the fourth period is the date after the current date, and the period duration of the fourth period is the same as the period duration of the third period. In a second aspect, embodiments of the present disclosure provide a method for training a conversion rate prediction model, the method including: Acquiring sample user information of a sample user, acquiring a conversion state label of the sample user in a first sample period, and determining a sample target crowd to which the sample user belongs based on the sample user information; taking the sample date as a first sample end date of a second sample period, acquiring sample time sequence data of a sample target crowd for executing preset behaviors in the second sample period, wherein the sample date is a previous date of the first sample start date of the first sample period; Determining historical sample time sequence data corresponding to a plurality of third sample periods in a crowd historical behavior time sequence library corresponding to sample target crowds, determining target historical sample time sequence data matched with the sample time sequence data from the historical sample time sequence data, wherein the period duration of the third sample periods is the same as the period duration of the second sample periods before the second sample periods; Determining a second sample end date of the target historical sample time sequence data, taking the date after the second sample end date as a second sample starting date of a fourth sample period, and obtaining sample conversion rate reference time sequence data in the fourth sample period; Constructing a training sample set based on the conversion state label, the sample conversion rate reference time sequence data and the sample user information; inputting a target training sample into a conversion rate prediction model t