Method and device for predicting transformation probability

Abstract

The invention provides a transformation probability prediction method and device, and the method comprises the steps: obtaining target variables corresponding to a plurality of historical users of at least one historical period and actual features under each feature type; performing binning processing on the plurality of actual features under each feature type according to a preset binning number and a target variable to obtain a binning feature under each feature type; screening out at least one target feature type from the plurality of feature types according to the binning feature corresponding to each feature type; determining a plurality of feature classification groups by calculating the Cartesian product of the binning features corresponding to each target feature type; calculating a predicted transformation probability of each feature classification group according to the actual transformation probability corresponding to each feature classification group in a plurality of historical periods of the target course; and according to the current user number and the predicted conversion probability corresponding to each feature classification group, calculating a predicted course conversion probability for the target course in the current period.

Description

Technical Field

[0001] The present application relates to the field of data analysis technology, and in particular to a method and device for predicting conversion probability. Background Art

[0002] In the prior art, time series analysis methods are used to predict the trend of conversion probabilities of online courses of different classes and sessions from trial listening to formal subscriptions. Time series analysis can only reflect the conversion probability of each period, but cannot analyze which groups of people only listen to the course and which users are likely to convert from trial listening to formal subscriptions. Subsequently, targeted online course recommendations cannot be made for groups that are likely to convert to formal subscriptions, which affects the prediction of conversion probability. Summary of the invention

[0003] In view of this, the purpose of the present application is to at least provide a method and device for predicting conversion probability, by binning the actual features of multiple historical users under each feature category, so as to bin each feature category into a preset number of bins, and screen out at least one target feature category from multiple feature categories in combination with the target variable of each historical user and the binning result, and then calculate the Cartesian product of the binned features of each target feature category to determine multiple feature classification groups, and then, according to the actual conversion probability corresponding to the target course of each feature classification group in the historical period, solve the predicted conversion probability of each feature classification group in the current period, so as to calculate the predicted course conversion probability of the target course in the current period through the current number of users corresponding to each feature classification group and the predicted conversion probability, thereby solving the technical problem that the conversion probability cannot be accurately predicted in the prior art and achieving the technical effect of improving the accuracy of conversion probability prediction.

[0004] This application mainly includes the following aspects:

[0005] In a first aspect, an embodiment of the present application provides a method for predicting conversion probability, the method comprising: obtaining target variables corresponding to multiple historical users of at least one historical period and actual features under each feature category, the target variable being used to describe whether the historical user has converted from a trial to a formal subscription to a target course of the historical period; binning multiple actual features under each feature category according to a preset number of bins and the target variable to obtain bin features under each feature category; screening out at least one target feature category from multiple feature categories based on the bin features corresponding to each feature category; determining multiple feature classification groups by calculating the Cartesian product of the bin features corresponding to each target feature category; calculating the predicted conversion probability of each feature classification group based on the actual conversion probabilities corresponding to each feature classification group in multiple historical periods of the target course, the conversion probability being used to describe the probability of converting from a trial to a formal subscription to the target course of the current period; calculating the predicted course conversion probability for the target course of the current period based on the number of current users corresponding to each feature classification group and the predicted conversion probability.

[0006] Optionally, one historical user corresponds to one actual feature under one feature category, and multiple actual features under each feature category are binned in the following manner: when the number of multiple actual features for multiple historical users under this feature category is greater than a preset bin number, any actual features are merged to obtain multiple actual feature groups, one actual feature group includes at least one merged actual feature, and the number of features in one actual feature group is the preset bin number; based on the target variables of each historical user corresponding to the merged actual feature group, the merged chi-square value of this feature category for each actual feature group is calculated; and each actual feature in the actual feature group corresponding to the smallest merged chi-square value is used as the bin feature of this feature category.

[0007] Optionally, historical users include trial users who have only tried the target courses for historical sessions and formal subscribers who have converted from trial to formal subscription to the target courses for historical sessions, and at least one target feature category is screened out from multiple feature categories in the following manner: for each feature category, based on the number of users of the trial users and the number of users of the formal subscribers corresponding to each of the boxed features under the feature category, the actual information value of the feature category is calculated; based on the time sequence corresponding to the multiple historical users, the multiple historical users are divided into a first user set and a second user set, and then based on the number of first users and the number of second users corresponding to each of the boxed features of each feature category, the actual group stability index of each feature category is calculated; for each feature category, based on the comparison result of the actual group stability index of the feature category with the preset group stability index and the comparison result of the actual information value with the preset information value, it is determined whether the feature category is the target feature category.

[0008] Optionally, the predicted conversion probability of each feature classification group is calculated based on the actual conversion probability corresponding to each feature classification group in multiple historical periods of the target course, including: for each feature classification group, calculating the product of the actual conversion probability corresponding to the feature classification group in each historical period and a preset influence weight, and taking the sum of the products corresponding to each historical period as the predicted conversion probability of the feature classification group for the target course of the current period.

[0009] Optionally, the predicted course conversion probability for the target course of the current session is calculated in the following manner: the current user proportion of each feature classification group is calculated; based on the predicted conversion probability corresponding to each feature classification group and the current user proportion, the predicted course conversion probability is calculated.

[0010] Optionally, the method also includes: determining the change in the proportion of users of each feature classification group in the current period; calculating the conversion contribution corresponding to each feature classification group based on the predicted course conversion probability and the change in the proportion of users corresponding to each feature classification group; and analyzing the impact of each feature classification group on the predicted course conversion probability by sorting the conversion contribution corresponding to each feature classification group.

[0011] Optionally, determining the change in user share of each feature classification group in the current period includes: calculating the average historical user share of each feature classification group in multiple historical periods, and calculating the current user share of each feature classification group; determining the change in user share corresponding to each feature classification group based on the current user share and the average historical user share corresponding to each feature classification group.

[0012] In a second aspect, an embodiment of the present application further provides a prediction device for conversion probability, the device comprising: an acquisition module, for acquiring target variables corresponding to multiple historical users of at least one historical period and actual features under each feature category, the target variable being used to describe whether the historical user has converted from a trial to a formal subscription to a target course of the historical period; a binning module, for binning multiple actual features under each feature category according to a preset number of bins and the target variable, so as to obtain bin features under each feature category; a screening module, for screening at least one target feature category from multiple feature categories based on the bin features corresponding to each feature category; a determination module, for determining multiple feature classification groups by calculating the Cartesian product of the bin features corresponding to each target feature category; a first calculation module, for calculating the predicted conversion probability of each feature classification group based on the actual conversion probabilities corresponding to each feature classification group in multiple historical periods of the target course, the conversion probability being used to describe the probability of converting from a trial to a formal subscription to the target course of the current period; a second calculation module, for calculating the predicted course conversion probability for the target course of the current period based on the number of current users corresponding to each feature classification group and the predicted conversion probability.

[0013] In a third aspect, an embodiment of the present application further provides an electronic device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to execute the steps of the method for predicting conversion probability described in the first aspect or any possible implementation manner of the first aspect.

[0014] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for predicting conversion probability described in the first aspect or any possible implementation manner of the first aspect are executed.

[0015] A method and device for predicting conversion probability provided by an embodiment of the present application include: obtaining target variables corresponding to multiple historical users of at least one historical period and actual features under each feature category, wherein the target variable is used to describe whether the historical user has converted from a trial to a formal subscription to a target course of the historical period; binning multiple actual features under each feature category according to a preset number of bins and the target variable to obtain bin features under each feature category; screening out at least one target feature category from multiple feature categories based on the bin features corresponding to each feature category; determining multiple feature classification groups by calculating the Cartesian product of the bin features corresponding to each target feature category; calculating a predicted conversion probability for each feature classification group based on the actual conversion probabilities corresponding to each feature classification group in multiple historical periods of the target course, wherein the conversion probability is used to describe the probability of converting from a trial to a formal subscription to the target course of the current period; calculating a predicted course conversion probability for the target course of the current period based on the number of current users corresponding to each feature classification group and the predicted conversion probability. The actual features of multiple historical users under each feature category are binned to a preset number of bins, and at least one target feature category is screened out from multiple feature categories in combination with the target variables and binning results of each historical user. The Cartesian product of the binned features of each target feature category is then calculated to determine multiple feature classification groups. Furthermore, the predicted conversion probability of each feature classification group in the current period is solved according to the actual conversion probability corresponding to the target course of each feature classification group in the historical period. The predicted course conversion probability of the target course in the current period is calculated through the current number of users and the predicted conversion probability corresponding to each feature classification group. This solves the technical problem of the failure to accurately predict the conversion probability in the prior art and achieves the technical effect of improving the accuracy of conversion probability prediction.

[0016] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A flow chart of a method for predicting conversion probability provided in an embodiment of the present application is shown.

[0019] Figure 2A functional module diagram of a conversion probability prediction device provided in an embodiment of the present application is shown.

[0020] Figure 3 A schematic structural diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0021] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art, under the guidance of the content of the present application, can add one or more other operations to the flowchart, or remove one or more operations from the flowchart.

[0022] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0023] In the prior art, trend changes are generally predicted through time series analysis, regression analysis, machine learning methods, etc. Different methods have different degrees of feature utilization and complexity of the reasoning process. At present, time series analysis is generally used to predict the conversion probability of online courses of different classes and sessions. The conversion probability refers to the probability that a user converts from a trial to a formal subscription course. This method adopts an autoregressive approach, does not consider user characteristics, and mainly adopts the idea of ​​weighted average, which is simple and easy to understand. However, time series analysis mainly captures periodic trends and does not consider the impact of users with different characteristics on the conversion probability, resulting in a certain difference between the subsequent actual conversion probability and the predicted conversion probability. In addition, it is impossible to effectively analyze the structural characteristics of the population, and it is impossible to conduct efficient attribution analysis to clarify the reasons for the large difference between the predicted conversion probability and the actual conversion probability.

[0024] Based on this, the embodiment of the present application provides a method and device for predicting conversion probability, which bins the actual features of multiple historical users under each feature category to bin each feature category into a preset number of bins, and selects at least one target feature category from multiple feature categories in combination with the target variable and binning result of each historical user, and then calculates the Cartesian product of the binned features of each target feature category to determine multiple feature classification groups, and then solves the predicted conversion probability of each feature classification group in the current period according to the actual conversion probability corresponding to the target course of each feature classification group in the historical period, so as to calculate the predicted course conversion probability of the target course in the current period through the current number of users and the predicted conversion probability corresponding to each feature classification group, thereby solving the technical problem of the failure to accurately predict the conversion probability in the prior art and achieving the technical effect of improving the accuracy of conversion probability prediction, which is specifically as follows:

[0025] See also Figure 1 , Figure 1 This is a flow chart of a method for predicting conversion probability provided by an embodiment of the present application. Figure 1 As shown, the method for predicting conversion probability provided by the embodiment of the present application includes the following steps:

[0026] S101: Obtain target variables corresponding to multiple historical users of at least one historical period and actual features under each feature type.

[0027] The target variable is used to describe whether the historical user has converted from a trial to a formal subscription to the target course of the historical session. That is, the multiple historical users of at least one historical session refer to students who study the target course of the corresponding historical session, including trial users who only listen to the target course of the historical session and formal subscribers who convert from a trial to a formal subscription to the target course of the historical session.

[0028] The target course can be an offline course or an online course. For example, online course software installed on the terminal device can be used to implement operations such as booking teaching materials for offline courses or online courses, viewing class notes, viewing and submitting class assignments, registering for trial classes, converting trial classes to formal subscriptions, and paying fees. A classroom page can also be provided for online courses.

[0029] That is to say, multiple historical users are users who have tried the target course of their corresponding historical sessions, among which some of the multiple historical users did not formally subscribe to the target course after the trial, and the other users formally subscribed to the target course after the trial. This application is to predict the conversion probability from trial to formal subscription, and further, it is still unknown how many users among the multiple users corresponding to the current session actually converted from trial to formal subscription.

[0030] Moreover, for each historical user, the historical user corresponds to multiple feature categories and corresponds to an actual feature under each feature category. The actual feature refers to the actual feature content of the historical user under the feature category. Exemplarily, multiple feature categories include the city to which they belong, the recommended channel, the user's grade, the user's parents' income level, etc. Furthermore, the actual features of user A under each feature category include: the city to which they belong is Beijing, the recommended channel is live broadcast, the user's grade is fourth grade, and the user's parents' income level is well-off. Furthermore, the actual feature corresponding to each historical user under each feature category can be obtained by the user filling out a questionnaire or the like.

[0031] Among them, the actual characteristics corresponding to the recommendation channels include: live broadcast, promotional web pages, offline sales, recommendations from other users, etc.

[0032] S102: binning multiple actual features under each feature category according to a preset bin number and the target variable to obtain bin features under each feature category.

[0033] Specifically, multiple actual features under each feature type are binned in the following manner: when the number of multiple actual features for multiple historical users under the feature type is greater than the preset bin number, any actual features are merged to obtain multiple actual feature groups, one actual feature group includes at least one merged actual feature, and the number of features in one actual feature group is the preset bin number; based on the target variables of each historical user corresponding to the merged actual feature group, the merged chi-square value of the feature type for each actual feature group is calculated; and each actual feature in the actual feature group corresponding to the smallest merged chi-square value is used as the bin feature of the feature type.

[0034] That is to say, for multiple historical users, one feature type corresponds to at least one actual feature, and the preset number of bins is a positive integer greater than 1. If the number of at least one actual feature corresponding to a feature type is less than or equal to the preset number of bins, there is no need to bin each actual feature of the feature type; if the number of at least one actual feature corresponding to a feature type is greater than the preset number of bins, multiple actual features of the feature type need to be binned.

[0035] The combined chi-square value of the combined actual feature group is used to express the degree of association between each feature under the actual feature group and the target variable (i.e., whether it is converted into a formal subscription). The smallest combined chi-square value is used to express that each feature under the corresponding actual feature group is similar in affecting the target variable, that is, each feature under the actual feature group corresponding to the smallest combined chi-square value can be understood as a class of features under its feature type. The preset number of bins is generally set to 3.

[0036] Specifically, the combined multiple actual feature groups are obtained in the following manner: for each feature type, the number of actual features n of the feature type is N, then by performing a combination of the actual features n1, n2, ..., n N Perform random merging so that the number of features after merging is the preset number of bins I.

[0037] For example, if the actual features of the cities are city a, city b, city c, and city d, and the preset number of bins is 3, then the actual feature groups after merging are as follows: actual feature group 1 is city a+city b, city c, and city d, actual feature group 2 is city a+city c, city b, and city d, actual feature group 3 is city a+city d, city b, and city c, actual feature group 4 is city b+city c, city a, and city d, actual feature group 5 is city b+city d, city a, and city c, and actual feature group 6 is city c+city d, city a, and city b. Furthermore, each actual feature group includes at least one merged actual feature, that is, city a+city b in actual feature group 1 is used as a merged actual feature, and the actual features of actual feature group 1 after merging are city a+city b, city c, and city d.

[0038] Specifically, the merged chi-square value is calculated through the following steps: construct a merged cross-tabulation, in which the rows represent the actual features corresponding to the feature type after the merger, and the columns represent the historical users as trial users or formal subscribers, so that each cell in the table represents the number of trial users or formal subscribers corresponding to the merged actual feature; calculate the expected frequency of each merged actual feature for trial users or formal subscribers according to the merged cross-tabulation; calculate the merged chi-square value of the feature type based on the number of trial users or formal subscribers corresponding to each merged actual feature and the expected frequency.

[0039] Specifically, the combined chi-square value is calculated using the following formula:

[0040]

[0041] In formula (1), x 2 Refers to the combined chi-square value of an actual feature group after merging, i refers to the actual feature i of an actual feature group after merging; I refers to the preset number of bins, that is, the actual number of features of an actual feature group; j1 refers to the trial user; j2 refers to the formal subscription user; It refers to the number of audition users corresponding to the actual feature i of an actual feature group; It refers to the number of formal subscribers corresponding to the actual feature i of an actual feature group; It refers to the expected frequency of audition users corresponding to the actual feature i of an actual feature group; It refers to the expected frequency of formal subscribed users corresponding to the actual feature i of an actual feature group.

[0042] Specifically, the expected frequency is calculated by the following formula:

[0043]

[0044] In formula (2), E i,j Refers to the expected frequency of trial listening users or formal subscribing users corresponding to the actual feature i of an actual feature group; R i Refers to the number of users with actual feature i in an actual feature group; C j It refers to the number of trial users or formal subscription users among multiple historical users; M refers to the total number of multiple historical users in at least one historical period, that is, the number of multiple historical users.

[0045] Exemplarily, according to the actual features of actual feature group 1 including city a+city b, city c and city d, the expected frequencies corresponding to trial users and formal subscribers in city a+city b, the expected frequencies corresponding to trial users and formal subscribers in city c, and the expected frequencies corresponding to trial users and formal subscribers in city d are calculated according to formula (2). Then, in combination with formula (1), the chi-square values ​​corresponding to trial users and formal subscribers in city a+city b, the chi-square values ​​corresponding to trial users and formal subscribers in city c, and the chi-square values ​​corresponding to trial users and formal subscribers in city d are summed to obtain the combined chi-square value of actual feature group 1.

[0046] Exemplarily, for each actual feature group, each actual feature in the actual feature group corresponding to the smallest combined chi-square value is used as a binning feature. If each actual feature of actual feature group 2 is city a+city c, city b, and city d, and actual feature group 2 is the actual feature group corresponding to the smallest combined chi-square value, then city a+city c, city b, and city d are used as binning features. In other words, city a+city c means that the city a or the city c is used as a binning feature; the historical users corresponding to the city a or the historical users corresponding to the city c are used as the historical users corresponding to the binning feature of city a+city c. That is to say, city a and city c have similar influences on the target variable. Therefore, city a and city c are used as a binning feature. Thus, in the feature type of city, the actual features (city a, city b, city c, city d) are simplified to binning features (city a + city c, city b, city d) to reduce the number of features under one feature type, thereby simplifying the number of feature classification groups obtained by the subsequent Cartesian product and simplifying subsequent calculations.

[0047] For example, after determining the binning characteristics of the recommendation channel, the recommendation channel with the lowest cost can be selected as the recommendation channel for actual application according to the application costs of multiple actual characteristics with similar influence on the target variable determined after binning. For example, if it is determined after binning that the influence of live broadcast and offline recommendation on the target variable is similar, the application cost corresponding to live broadcast and offline recommendation can be referred to and the method with the lowest application cost can be selected as the recommendation channel for actual application, so as to achieve the technical effect of reducing the recommendation cost.

[0048] For example, the binning features for recommendation channels are live broadcast + promotional webpage, offline sales, and recommendations from other users. The promotional webpage can be a recommended channel for converting a trial class for the user into a formal subscription that pops up every time the user logs into the online course software. The live broadcast can be a recommended channel for converting a trial class for the user into a formal subscription by popping up on the classroom page at the end of the live course of the online course. Furthermore, since the influence of live broadcast and promotional webpage on the target variable is similar, push messages can be sent to a user only through live broadcast or promotional webpage, which reduces server processing steps, reduces resource usage, and does not significantly change the push success rate.

[0049] S103: Screening out at least one target feature category from a plurality of feature categories according to the binning features corresponding to each feature category.

[0050] Specifically, at least one target feature type is screened out from multiple feature types in the following manner: for each feature type, based on the number of trial users and the number of formal subscribed users corresponding to each of the boxed features under the feature type, the actual information value of the feature type is calculated; based on the time sequence corresponding to multiple historical users, multiple historical users are split into a first user set and a second user set, and then based on the number of first users and the number of second users corresponding to each of the boxed features of each feature type, the actual group stability index of each feature type is calculated; for each feature type, based on the comparison result of the actual group stability index of the feature type with the preset group stability index and the comparison result of the actual information value with the preset information value, it is determined whether the feature type is a target feature type.

[0051] That is to say, after each feature type is binned, the number of binned features corresponding to each feature type is less than or equal to the preset bin number. In this way, the PSI (population stability index) and IV (information value) corresponding to each feature type after binning are calculated, and each feature type is screened in combination with the preset population stability index and the preset information value to screen out at least one target feature type.

[0052] Among them, the larger the IV, the stronger the predictive ability of the feature type in determining whether the user is a formal subscriber; the smaller the PSI, the better the stability of the feature type.

[0053] Specifically, the multiple historical users are divided into the first user set and the second user set in the following manner: the multiple historical users are sorted according to the time when each historical user registered for the target course, and the multiple historical users are divided into the first user set and the second user set according to the sorting result. Exemplarily, according to the time sequence of registration for the target course, a part of the historical users who registered first are taken as the first user set, and the remaining users are taken as the second user set, so that the number of users in the first user set is the same as that in the second user set, or the difference in the number of users in the first user set and the second user set meets the preset difference range.

[0054] The first number of users corresponding to each binning feature refers to the number of users in the first user set that belong to the binning feature, and the second number of users corresponding to each binning feature refers to the number of users in the second user set that belong to the binning feature.

[0055] Specifically, the actual group stability index is calculated by the following formula:

[0056]

[0057] In formula (3), PSI p Refers to the actual population stability index of the pth feature category; i refers to the bin feature i of a feature category after binning; I refers to the total number of features of the feature category after binning, that is, the preset number of bins; A i It refers to the ratio of the number of first users corresponding to the bin feature i to the number of users of multiple historical users; A′ i It refers to the ratio of the number of second users corresponding to the bin feature i to the number of users of multiple users.

[0058] Specifically, the actual information value is calculated by the following formula:

[0059]

[0060] In formula (4), IV p Refers to the actual information value of the pth feature category; i refers to the binned feature i of a feature category after binning; I refers to the total number of features of a feature category after binning, which is generally the preset number of bins; B i It refers to the ratio of the number of formal subscribed users corresponding to bin feature i to the number of multiple historical users; B′ i It refers to the ratio of the number of audition users corresponding to bin feature i to the number of multiple historical users.

[0061] That is to say, the actual population stability index corresponding to each feature type is calculated by formula (3), and the actual information value corresponding to each feature type is calculated by formula (4); for each feature type, when the actual population stability index of the feature type is less than the preset population stability index and the actual information value is greater than the preset information value, the feature type is used as the target feature type. Furthermore, by calculating the actual population stability index and the actual information value, the target feature types with accurate prediction and good stability are screened out. While screening out feature types with better prediction effect and stability, it can also play a role in reducing the number of feature types, thereby reducing the amount of subsequent calculations. For example, if the influence of the feature type on the predicted conversion probability is low (i.e., IV is less than or equal to the preset information value) and / or the stability is poor (i.e., PSI is greater than or equal to the preset population stability index), the feature type is removed.

[0062] For example, the preset group stability index is generally 0.1, and the preset information value is generally 0.02. Therefore, when the actual group stability index and the actual information value meet the corresponding requirements, it indicates that the feature type has the ability to influence the prediction of the user as a formal subscription user and has stability.

[0063] S104: Determine multiple feature classification groups by calculating the Cartesian product of the binned features corresponding to each target feature type.

[0064] For example, if the target feature types are cities and recommended channels, the binning features of the cities are city a+city c, city b, and city d, and the binning features of the recommended channels are live broadcasting + offline sales, promotional webpages, and recommendations from other users. Furthermore, the feature classification groups calculated by the Cartesian product include: (city a+city c, live broadcasting + offline sales), (city a+city c, promotional webpages), (city a+city c, recommendations from other users), (city b, live broadcasting + offline sales), (city b, promotional webpages), (city b, recommendations from other users), (city d, live broadcasting + offline sales), (city d, promotional webpages), and (city d, recommendations from other users).

[0065] That is to say, a boxed feature is selected in each target feature type to obtain multiple non-repetitive feature classification groups, so as to classify multiple historical users according to the boxed features under each target feature type, so as to facilitate the subsequent calculation of the predicted conversion probability of each feature classification group and the predicted course conversion probability of the target course, as well as the subsequent analysis of the impact of each feature classification group on the predicted course conversion probability of the target course.

[0066] S105: Calculate the predicted conversion probability of each feature classification group based on the actual conversion probability of each feature classification group corresponding to multiple historical sessions of the target course.

[0067] The conversion probability is used to describe the probability of converting from trial listening to formal subscription of the target course of the current term. In other words, the predicted conversion probability refers to the prediction of the probability of converting from trial listening to formal subscription among multiple current users corresponding to the target course of the current term.

[0068] Among them, the at least one historical period mentioned in the foregoing is to determine the feature classification group, and the multiple historical periods are to calculate the actual conversion probability. Since the calculation of the feature classification group and the actual conversion probability do not conflict, the multiple historical periods and the at least one historical period can be the same historical period or different historical periods. Exemplarily, the multiple historical periods can be the previous period, the previous two periods, and the previous three periods of the current period; the at least one historical period can be at least one period before the current period, and can be the previous period of the current period.

[0069] The method calculates the predicted conversion probability of each feature classification group based on the actual conversion probability corresponding to each feature classification group in multiple historical periods of the target course, including: for each feature classification group, calculating the product of the actual conversion probability corresponding to the feature classification group in each historical period and the preset influence weight, and taking the sum of the products corresponding to each historical period as the predicted conversion probability of the feature classification group for the target course of the current period.

[0070] Among them, the preset influence weight is used to indicate the degree of influence of the historical period on the time equivalent to the current period. For example, as parents pay more attention to education, the number of users who convert from trial to formal subscription will increase. Therefore, the actual conversion probability of the historical period closer to the current period will have a greater impact on the conversion probability of the current period, and the preset influence weight will also be greater.

[0071] Among them, the preset influence weights corresponding to each historical period need to be normalized, and the weights corresponding to each historical period can be determined by actual experience or analysis by the experimenter. Exemplarily, each historical period is the previous period, the previous two periods, and the previous three periods of the current period, among which, according to actual experience, the weight corresponding to the previous period is set to 0.8, the weight corresponding to the previous two periods is set to 0.15, and the weight corresponding to the previous three periods is set to 0.05.

[0072] Exemplarily, the predicted conversion probability of each feature classification group is calculated by the following formula:

[0073]

[0074] In formula (5), y q refers to the predicted conversion probability of the qth feature classification group; wt Refers to the weights corresponding to the first t periods after normalization, that is, w1 is the weight corresponding to the first period, w2 is the weight corresponding to the first two periods, and w3 is the weight corresponding to the first three periods; T refers to the maximum value of t in the first t periods; y t,q Refers to the actual conversion probability of the qth feature classification group in the first t periods.

[0075] Among them, y t,q It is calculated in the following way: for each feature classification group, determine the number of users and formally subscribed users corresponding to the feature classification group in the first t periods, and use the ratio of the formally subscribed users corresponding to the feature classification group in the first t periods to its number of users as the actual conversion probability of the feature classification group in the first t periods.

[0076] The predicted conversion probability of each feature classification group indicates the predicted probability of the current user corresponding to each feature classification group converting from trial listening to formal subscription. Furthermore, the predicted conversion probability of each feature classification group is calculated by referring to the conversion status of historical sessions and the influencing factors of the time when the historical sessions are equivalent to the current session, so as to increase the prediction accuracy of the predicted conversion probability of each feature classification group.

[0077] Exemplarily, if a feature classification group is (city a + city c, live broadcast + offline sales), the number of users whose city is city a or city c and whose recommended channel is live broadcast or offline sales in the previous period of this feature classification group is determined, and the actual formal subscription users are determined from the number of users, and the ratio of the formal subscription users to the number of users in the previous period is taken as the actual conversion probability of this feature classification group in the previous period.

[0078] S106: Calculate the predicted course conversion probability for the target course in the current session based on the current number of users corresponding to each feature classification group and the predicted conversion probability.

[0079] The predicted course conversion probability for the target course of the current session is calculated in the following manner: the current user ratio of each feature classification group is calculated; based on the predicted conversion probability corresponding to each feature classification group and the current user ratio, the predicted course conversion probability is calculated.

[0080] Among them, the current user ratio of each feature classification group is calculated in the following way: multiple current users corresponding to the target course of the current period are obtained, and the number of current users corresponding to each feature classification group is determined; for each feature classification group, the ratio of the current user number of the feature classification group to the number of users of the multiple current users is used as the current user ratio of the feature classification group.

[0081] Among them, the multiple current users corresponding to the target course of the current term have not yet been distinguished as trial users or those converted from trial users to formal subscription users. In other words, the multiple current users corresponding to the target course of the current term can all be understood as trial users, and further, the predicted course conversion probability can be understood as predicting the course conversion probability of the target course actually being converted into a formal subscription in the current term before it is actually converted into a formal subscription.

[0082] For example, if a feature classification group is (city a + city c, live broadcast + offline sales), the ratio of the number of current users who are recommended through live broadcast or offline sales in city a or city c to the number of multiple current users who have signed up for the target course in the current session will be taken as the current user ratio of the feature classification group.

[0083] Specifically, the predicted course conversion probability is calculated by the following formula:

[0084]

[0085] In formula (6), y refers to the predicted course conversion probability of the target course in the current period; y q refers to the predicted conversion probability of the qth feature classification group; p q It refers to the current user proportion corresponding to the qth feature classification group, that is, the user proportion corresponding to the qth feature classification group in the current period; Q refers to the total number of feature classification groups.

[0086] Furthermore, the product of the current user proportion of each feature classification group and its corresponding predicted conversion probability is summed to obtain the predicted course conversion probability for the target course in the current session. The predicted course conversion probability represents the probability of a user converting from a trial to a formal subscription to the target course in the current session.

[0087] Furthermore, since the calculation method of the predicted course conversion probability takes into account the current user proportion of each feature classification group and its predicted conversion probability, the current users are classified from the perspective of the feature classification group, thereby taking into account the population size of different types of people and the impact of the predicted conversion probability on the predicted course conversion probability.

[0088] For example, cities that place more emphasis on education will have a greater probability of converting from trial lessons to formal subscriptions. Students in higher grades may also take fewer courses other than subject-based courses to reduce the probability of converting from trial lessons to formal subscriptions. Therefore, considering different groups of people can increase the accuracy of predicting the probability of course conversion.

[0089] The method also includes: determining the change in the user share of each feature classification group in the current period; calculating the conversion contribution corresponding to each feature classification group based on the predicted course conversion probability and the change in the user share corresponding to each feature classification group; and analyzing the impact of each feature classification group on the predicted course conversion probability by sorting the conversion contribution corresponding to each feature classification group.

[0090] The method of determining the change in the user share of each feature classification group in the current period includes: calculating the average historical user share of each feature classification group in multiple historical periods, and calculating the current user share of each feature classification group; and determining the change in the user share corresponding to each feature classification group based on the current user share and the average historical user share corresponding to each feature classification group.

[0091] That is to say, for each feature classification group, the ratio of the number of users of the feature classification group in the current period to the number of users of multiple current users is used as the current user ratio of the feature classification group; the number of users corresponding to the feature classification group in multiple historical periods is calculated and compared with the number of historical users in each historical period, and the ratio is used as the historical user ratio of each historical period, and the average value of the historical user ratio of each historical period is calculated; the difference between the current user ratio and the average value is used as the change in the user ratio of the current period.

[0092] Among them, the change in the user proportion of the current period reflects the change in the user proportion of each feature classification group in the current period compared with multiple historical periods. For each feature classification group, if the change in the user proportion of the feature classification group is a positive number, it means that the number of users corresponding to the feature classification group in the current period is greater than the number of users in the historical period; if the change in the user proportion of the feature classification group is a negative number, it means that the number of users corresponding to the feature classification group in the current period is less than the number of users in the historical period.

[0093] For each feature classification group, the product of the predicted course conversion probability and the change in the user proportion corresponding to the feature classification group is taken as the conversion contribution corresponding to the feature classification group.

[0094] Among them, the product of the absolute value of the change in the user proportion corresponding to each feature classification group and the predicted course conversion probability can be used as the conversion contribution corresponding to each feature classification group, and then the conversion contributions corresponding to each feature classification group are arranged in descending order, so as to reflect which feature classification groups have a greater impact on the predicted course conversion probability in the current period.

[0095] In other words, the absolute value of the change in the user ratio is used to reflect the fluctuation range of each feature classification group in the current period relative to the historical period; the conversion contribution is used to characterize the degree of influence of each feature classification group on the predicted course conversion probability. Furthermore, by comparing the difference between the predicted course conversion probability of the current period and the actual course conversion probability of the historical period, and then combining the conversion contribution corresponding to each feature classification group, we can analyze which feature classification groups have a greater impact on the fluctuation of the predicted course conversion probability, so as to facilitate marketing personnel to achieve analysis and processing such as positioning and attribution of changes.

[0096] Exemplarily, the conversion contribution corresponding to each feature classification group is arranged in descending order to obtain feature classification groups sorted from large to small according to the conversion contribution. For feature classification groups with high rankings, the probability of conversion from trial listening to actual subscription can be increased by increasing the recommendation frequency for each user of the current period of the feature classification group, so as to increase the recommendation accuracy and improve the probability of successful recommendation; for feature classification groups with low rankings, the recommendation cost can be reduced by reducing the recommendation frequency for each user of the current period of the feature classification group.

[0097] Among them, the preset positions after sorting can be used to divide the feature classification groups corresponding to the front and back rankings, that is, the feature classification groups located at and before the preset positions after descending sorting are regarded as the feature classification groups with the front rankings, and the feature classification groups located after the preset positions after descending sorting are regarded as the feature classification groups with the back rankings, so as to realize targeted recommendations and further improve the accuracy of recommendations.

[0098] Based on the same application concept, the embodiments of the present application also provide a conversion probability prediction device corresponding to the conversion probability prediction method provided in the above embodiments. Since the principle of solving the problem by the device in the embodiments of the present application is similar to the conversion probability prediction method in the above embodiments of the present application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0099] like Figure 2 As shown, Figure 2A functional module diagram of a conversion probability prediction device provided in an embodiment of the present application. A conversion probability prediction device 10 includes: an acquisition module 101, used to obtain target variables corresponding to multiple historical users of at least one historical period and actual features under each feature category, wherein the target variable is used to describe whether the historical user has converted from a trial to a formal subscription to a target course of the historical period; a binning module 102, used to bin multiple actual features under each feature category according to a preset bin number and the target variable, so as to obtain bin features under each feature category; a screening module 103, used to screen out at least one target feature from multiple feature categories based on the bin features corresponding to each feature category. feature type; a determination module 104, used to determine multiple feature classification groups by calculating the Cartesian product of the binned features corresponding to each target feature type; a first calculation module 105, used to calculate the predicted conversion probability of each feature classification group according to the actual conversion probability corresponding to each feature classification group in multiple historical sessions of the target course, the conversion probability is used to describe the probability of converting from a trial to a formal subscription to the target course of the current session; a second calculation module 106, used to calculate the predicted course conversion probability for the target course of the current session according to the current number of users corresponding to each feature classification group and the predicted conversion probability.

[0100] Based on the same application concept, see Figure 3 As shown, it is a structural diagram of an electronic device provided in an embodiment of the present application, the electronic device 20 includes: a processor 201, a memory 202 and a bus 203, the memory 202 stores machine-readable instructions executable by the processor 201, when the electronic device 20 is running, the processor 201 communicates with the memory 202 through the bus 203, and the machine-readable instructions are executed by the processor 201 when running to perform the steps of the conversion probability prediction method as described in any of the above embodiments.

[0101] Specifically, when the machine-readable instructions are executed by the processor 201, the following processing can be performed: obtaining target variables corresponding to multiple historical users of at least one historical period and actual features under each feature category, the target variable being used to describe whether the historical user has converted from a trial to a formal subscription to a target course of the historical period; binning multiple actual features under each feature category according to a preset number of bins and the target variable to obtain bin features under each feature category; screening out at least one target feature category from multiple feature categories based on the bin features corresponding to each feature category; determining multiple feature classification groups by calculating the Cartesian product of the bin features corresponding to each target feature category; calculating the predicted conversion probability of each feature classification group based on the actual conversion probability corresponding to each feature classification group in multiple historical periods of the target course, the conversion probability being used to describe the probability of converting from a trial to a formal subscription to the target course of the current period; calculating the predicted course conversion probability for the target course of the current period based on the number of current users corresponding to each feature classification group and the predicted conversion probability.

[0102] Based on the same application concept, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the conversion probability prediction method provided in the above embodiment are executed.

[0103] Specifically, the storage medium can be a general storage medium, such as a mobile disk, a hard disk, etc. When the computer program on the storage medium is run, the above-mentioned conversion probability prediction method can be executed, by binning the actual features of multiple historical users under each feature category, so as to bin each feature category into a preset number of bins, and combine the target variables and binning results of each historical user to screen out at least one target feature category from multiple feature categories, and then calculate the Cartesian product of the binned features of each target feature category to determine multiple feature classification groups, and then, according to the actual conversion probability corresponding to the target course of each feature classification group in the historical period, the predicted conversion probability of each feature classification group in the current period is solved, so as to calculate the predicted course conversion probability of the target course in the current period through the current number of users corresponding to each feature classification group and the predicted conversion probability, thereby solving the technical problem of the failure to accurately predict the conversion probability in the prior art and achieving the technical effect of improving the accuracy of conversion probability prediction.

[0104] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0105] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0106] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0107] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0108] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for predicting conversion probability, characterized in that: The method comprises: Obtain target variables corresponding to multiple historical users of at least one historical session and actual features under each feature type, wherein the target variables are used to describe whether the historical users have converted from trial listening to formally subscribing to the target course of the historical session; According to the preset bin number and the target variable, multiple actual features under each feature category are binned to obtain bin features under each feature category; According to the binning features corresponding to each feature type, at least one target feature type is selected from multiple feature types; By calculating the Cartesian product of the binned features corresponding to each target feature type, multiple feature classification groups are determined; According to the actual conversion probabilities corresponding to each feature classification group in multiple historical sessions of the target course, the predicted conversion probability of each feature classification group is calculated, and the conversion probability is used to describe the probability of converting from trial listening to formal subscription of the target course of the current session; Based on the current number of users corresponding to each feature classification group and the predicted conversion probability, the predicted course conversion probability for the target course of the current session is calculated.

2. The method according to claim 1, characterized in that One historical user corresponds to one actual feature under one feature category. Multiple actual features under each feature category are binned in the following way: When the number of multiple actual features for multiple historical users under this feature type is greater than the preset number of bins, any actual features are merged to obtain multiple actual feature groups, one actual feature group includes at least one merged actual feature, and the number of features in one actual feature group is the preset number of bins; According to the target variables of each historical user corresponding to the merged actual feature group, the merged chi-square value of the feature type for each actual feature group is calculated; Each actual feature in the actual feature group corresponding to the smallest combined chi-square value is used as the binning feature of the feature type.

3. The method according to claim 1 or 2, characterized in that: The historical users include trial users who only try out the target courses of the historical sessions and formal subscribers who convert from trial to formal subscription of the target courses of the historical sessions. At least one target feature type is selected from multiple feature types in the following manner: For each feature type, based on the number of trial users and the number of formal subscribers corresponding to each of the binned features under the feature type, the actual information value of the feature type is calculated; According to the time sequence corresponding to the multiple historical users, the multiple historical users are divided into a first user set and a second user set, and then based on the number of first users and the number of second users corresponding to each binned feature of each feature type, the actual group stability index of each feature type is calculated; For each feature type, it is determined whether the feature type is a target feature type based on a comparison result between the actual group stability index of the feature type and a preset group stability index and a comparison result between the actual information value and a preset information value.

4. The method according to claim 1, characterized in that: The step of calculating the predicted conversion probability of each feature classification group according to the actual conversion probability corresponding to each feature classification group in multiple historical sessions of the target course includes: For each feature classification group, the product of the actual conversion probability corresponding to the feature classification group in each historical period and the preset influence weight is calculated, and the sum of the products corresponding to each historical period is used as the predicted conversion probability of the feature classification group for the target course in the current period.

5. The method according to claim 1, characterized in that The predicted course conversion probability for the target course of the current session is calculated in the following way: Calculate the current user percentage of each feature classification group; The predicted course conversion probability is calculated based on the predicted conversion probability corresponding to each feature classification group and the current user ratio.

6. The method according to claim 1, characterized in that The method further comprises: Determine the change in the user proportion of each feature classification group in the current period; Calculate the conversion contribution of each feature classification group based on the predicted course conversion probability and the change in the user proportion corresponding to each feature classification group; By ranking the conversion contribution corresponding to each feature classification group, the impact of each feature classification group on the predicted course conversion probability is analyzed.

7. The method according to claim 6, characterized in that Determining the change in the user proportion of each feature classification group in the current period includes: Calculate the average historical user share of each feature classification group in multiple historical periods, and calculate the current user share of each feature classification group; According to the current user proportion and the average historical user proportion corresponding to each feature classification group, the change in user proportion corresponding to each feature classification group is determined.

8. A device for predicting conversion probability, characterized in that: The device comprises: An acquisition module, used to acquire target variables corresponding to multiple historical users of at least one historical session and actual features under each feature type, wherein the target variables are used to describe whether the historical users have converted from trial listening to formally subscribing to the target course of the historical session; A binning module is used to bin multiple actual features under each feature category according to a preset binning number and the target variable to obtain binning features under each feature category; A screening module, used to screen out at least one target feature category from a plurality of feature categories according to the binning features corresponding to each feature category; A determination module, used to determine a plurality of feature classification groups by calculating the Cartesian product of the binned features corresponding to each target feature type; A first calculation module is used to calculate the predicted conversion probability of each feature classification group according to the actual conversion probability corresponding to each feature classification group in multiple historical sessions of the target course, wherein the conversion probability is used to describe the probability of converting from trial listening to formal subscription of the target course of the current session; The second calculation module is used to calculate the predicted course conversion probability for the target course of the current session based on the current number of users corresponding to each feature classification group and the predicted conversion probability.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to execute the steps of the conversion probability prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for predicting conversion probability according to any one of claims 1 to 7 are executed.

Images