Game payment prediction method, model training method and device
By combining the decision tree model and the recurrent neural network model, using local and global feature attention weights for weighted calculations, the inaccuracy problem caused by the simplicity of game paid prediction models in the prior art is solved, and the accuracy of prediction results is improved.
Patent Information
- Application Number
- CN202010180136.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-03-13
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2040-03-13
AI Technical Summary
In the prior art, the game paid prediction model is simple to construct, resulting in inaccurate prediction results.
Using a method combining decision tree model and recurrent neural network model, paid features are extracted through the decision tree model and feature weighting is used to predict using the recurrent neural network model, and local and global feature attention weights are used to predict.
This improves the accuracy of paid predictions in games, and makes the model pay more attention to features that have a greater impact on the output variables, thereby improving the accuracy of the prediction results.
Smart Images

Figure CN111408143B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data prediction technology, and in particular to a game payment prediction method, a model training method and a device. Background Art
[0002] Machine learning has many applications in daily life, one of which is predicting time series. For example, for promotions on specific dates, how can merchants predict sales in advance to arrange procurement and warehousing, promotional pricing, and optimize distribution plans? How can we predict the traffic volume at a specific time in the urban road network to plan traffic reasonably? In these practical scenarios, artificial intelligence methods are needed to predict future data changes based on historical data.
[0003] In the existing technology, a preset model can be used to predict the daily game payment situation in the future, and the prediction results can be provided to the decision-making level to help the later operation of the game.
[0004] However, in the prior art, the model for predicting game payment is relatively simple, and there is no difference between the historical features, which leads to the problem of inaccurate results of game payment prediction. Summary of the invention
[0005] The purpose of this application is to provide a game payment prediction method, model training method and device to address the deficiencies in the above-mentioned prior art, so as to solve the problem that the model construction in the prior art is relatively simple, resulting in inaccurate prediction.
[0006] To achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:
[0007] In a first aspect, an embodiment of the present application provides a game payment prediction method, which is applied to a payment prediction model, wherein the payment prediction model includes: a first decision tree model and a recurrent neural network model; the method includes:
[0008] According to the payment feature set of the target game, the first decision tree model is used to obtain a first predicted payment feature of the target game;
[0009] According to the payment feature set and the first predicted payment feature, the recurrent neural network model is used to obtain a second predicted payment feature of the target game; wherein the preset recurrent neural network model is used to determine the local feature attention weights of various features in the payment feature set and the first predicted payment feature, and according to the payment feature set, the first predicted payment feature and the local feature attention weights of various features, the second predicted payment feature is obtained;
[0010] The first predicted payment feature and the second predicted payment feature are weightedly calculated to obtain a payment prediction result of the target game.
[0011] Optionally, the payment feature set includes: at least one type of payment feature.
[0012] The local feature attention weights of each type of feature include: the local feature attention weight of the at least one type of payment feature, and the local feature attention weight of the first predicted payment feature.
[0013] Optionally, the payment feature set includes at least one of the following features: historical payment features, expected payment features in the future, and trend features corresponding to a preset prediction time step.
[0014] Optionally, before obtaining the first predicted payment feature of the target game by using the first decision tree model according to the payment feature set of the target game, the method further includes:
[0015] Sorting the payment features in the payment feature set according to time to obtain a plurality of feature sequences corresponding to time information, wherein each feature sequence corresponding to time information includes: at least one payment feature, and each payment feature corresponds to a feature type;
[0016] The payment feature set includes: feature sequences corresponding to the multiple time information.
[0017] Optionally, the method further comprises:
[0018] The recurrent neural network model is used to obtain the second predicted payment feature based on the payment feature set, the first predicted payment feature, the local feature attention weights of each type of feature, and the global time attention weights of the feature sequence corresponding to each time information.
[0019] Optionally, the payment prediction model includes: a second decision tree model, wherein the recurrent neural network model is used according to the payment feature set and the first predicted payment feature, including:
[0020] According to the payment feature set and the first predicted payment feature, the second decision tree model is used to obtain a third predicted payment feature of the target game;
[0021] According to the payment feature set and the third predicted payment feature, the recurrent neural network model is used to obtain the second predicted payment feature;
[0022] The step of performing weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain the payment prediction result of the target game includes:
[0023] The first predicted payment feature, the second predicted payment feature and the third predicted payment feature are weightedly calculated to obtain a payment prediction result of the target game.
[0024] Optionally, the second predicted payment feature includes: a predicted payment feature corresponding to at least one predicted time step;
[0025] The step of performing weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain the payment prediction result of the target game includes:
[0026] A weighted calculation is performed on the first predicted payment feature and the predicted payment feature corresponding to each predicted time step to obtain a payment prediction result corresponding to each predicted time step.
[0027] In a second aspect, another embodiment of the present application provides a method for training a payment prediction model, wherein the payment prediction model includes: a first decision tree model and a recurrent neural network model, and the model training method includes:
[0028] According to the training payment feature set of the target game, the first decision tree model is used to obtain a fourth predicted payment feature of the target game;
[0029] Optimizing the parameters of the first decision tree according to the verified payment feature set of the target game and the fourth predicted payment feature;
[0030] According to the predicted payment feature output by the optimized first decision tree and the training payment feature set, the recurrent neural network model is used to obtain a fifth predicted payment feature of the target game;
[0031] Optimizing the parameters of the recurrent neural network model according to the verification payment feature set and the fifth prediction payment feature;
[0032] A weighted calculation is performed on the predicted payment features output by the optimized first decision tree and the predicted payment features output by the optimized recurrent neural network model.
[0033] Optionally, the training payment feature set includes: at least one type of training payment feature.
[0034] The local feature attention weights of each type of feature include: the local feature attention weight of the at least one type of training payment feature, and the local feature attention weight of the fourth predicted payment feature.
[0035] Optionally, before obtaining the fourth predicted payment feature of the target game by using the first decision tree model according to the training payment feature set of the target game, the method further includes:
[0036] Sort the training payment features in the training payment feature set according to time to obtain a plurality of training feature sequences corresponding to time information, wherein each training feature sequence corresponding to time information includes at least one training payment feature, and each training payment feature corresponds to a feature type;
[0037] The training payment feature set includes: a training feature sequence corresponding to the multiple time information.
[0038] Optionally, the method further comprises:
[0039] The recurrent neural network model is used to obtain the fifth predicted payment feature according to the training payment feature set, the fourth predicted payment feature, the local feature attention weights of various features, and the global time attention weights of the feature sequence corresponding to each time information;
[0040] The parameters of the recurrent neural network model also include: the global temporal attention weight.
[0041] Optionally, the payment prediction model includes: a second decision tree model; the predicted payment feature outputted by the optimized first decision tree and the training payment feature set, using the recurrent neural network model to obtain a fifth predicted payment feature of the target game, including:
[0042] According to the predicted payment feature output by the optimized first decision tree and the training payment feature set, the second decision tree model is used to obtain a sixth predicted payment feature of the target game;
[0043] Optimizing the parameters of the second decision tree model according to the verification payment feature set and the sixth prediction payment feature;
[0044] According to the predicted payment feature output by the optimized second decision tree and the training payment feature set, the recurrent neural network model is adopted to obtain the fifth predicted payment feature.
[0045] Optionally, before obtaining the fourth predicted payment feature of the target game by using the first decision tree model according to the training payment feature set of the target game, the method further includes:
[0046] Obtaining a sample payment feature set of the target game;
[0047] According to a preset division rule, the sample payment feature set is divided into a plurality of feature sets, and the plurality of feature sets include: the training payment feature set and the verification payment feature set.
[0048] Optionally, according to a preset division rule, the present payment feature set is divided into multiple feature sets, including:
[0049] According to the time information of the payment features in the sample payment feature set and a preset division ratio, the sample payment feature information set is divided into the multiple feature sets.
[0050] Optionally, the plurality of feature sets further include: a test payment feature set;
[0051] According to the test payment feature set, using the optimized first decision tree, obtaining prediction features of the first decision tree;
[0052] According to the prediction features of the first decision tree and the test payment feature set, the optimized recurrent neural network model is used to obtain the prediction features of the recurrent neural network;
[0053] A weighted calculation is performed based on the prediction features of the first decision tree and the prediction features of the recurrent neural network to obtain a payment prediction result of the target game during the test process, and the payment prediction result during the test process includes: the prediction features after the weighted calculation.
[0054] In a third aspect, another embodiment of the present application provides a game payment prediction device, which is applied to a payment prediction model, wherein the payment prediction model includes: a first decision tree model and a recurrent neural network model; the device includes: a prediction module and a calculation module, wherein:
[0055] The prediction module is used to obtain a first predicted payment feature of the target game using the first decision tree model according to the payment feature set of the target game;
[0056] The prediction module is used to obtain the second predicted payment feature of the target game using the recurrent neural network model according to the payment feature set and the first predicted payment feature; wherein the preset recurrent neural network model is used to determine the local feature attention weights of various features in the payment feature set and the first predicted payment feature, and obtain the second predicted payment feature according to the payment feature set, the first predicted payment feature and the local feature attention weights of various features;
[0057] The calculation module is used to perform weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain a payment prediction result of the target game.
[0058] Optionally, the device further comprises: a sorting module, configured to sort each payment feature in the payment feature set according to time, to obtain a plurality of feature sequences corresponding to time information, wherein each feature sequence corresponding to time information comprises: at least one payment feature, and each payment feature corresponds to a feature type;
[0059] The payment feature set includes: feature sequences corresponding to the multiple time information.
[0060] Optionally, the prediction module is also used to obtain the second predicted payment feature based on the payment feature set, the first predicted payment feature, the local feature attention weights of each type of feature, and the global time attention weight of the feature sequence corresponding to each time information.
[0061] Optionally, the prediction module is further used to obtain a third predicted payment feature of the target game using the second decision tree model according to the payment feature set and the first predicted payment feature;
[0062] The prediction module is further used to obtain the second predicted payment feature by using the recurrent neural network model according to the payment feature set and the third predicted payment feature;
[0063] The calculation module is further used to perform weighted calculation on the first predicted payment feature, the second predicted payment feature and the third predicted payment feature to obtain a payment prediction result of the target game.
[0064] Optionally, the second predicted payment feature includes: a predicted payment feature corresponding to at least one predicted time step;
[0065] The calculation module is further used to perform weighted calculation on the first predicted payment feature and the predicted payment feature corresponding to each prediction time step to obtain the payment prediction result corresponding to each prediction time step.
[0066] In a fourth aspect, another embodiment of the present application further provides a training device for a payment prediction model, wherein the payment prediction model includes: a first decision tree model and a recurrent neural network model, and the device includes: a prediction module, an optimization module and a calculation module, wherein:
[0067] The prediction module is used to obtain a fourth predicted payment feature of the target game using the first decision tree model according to the training payment feature set of the target game;
[0068] The optimization module is used to optimize the parameters of the first decision tree according to the verification payment feature set of the target game and the fourth predicted payment feature;
[0069] The prediction module is further used to obtain a fifth predicted payment feature of the target game using the recurrent neural network model according to the predicted payment feature output by the optimized first decision tree and the training payment feature set;
[0070] The optimization module is further used to optimize the parameters of the recurrent neural network model according to the verification payment feature set and the fifth prediction payment feature;
[0071] The calculation module is used to perform weighted calculation on the predicted payment features output by the optimized first decision tree and the predicted payment features output by the optimized recurrent neural network model.
[0072] Optionally, the device further comprises a sorting module, which is used to sort the training payment features in the training payment feature set according to time to obtain a plurality of training feature sequences corresponding to time information, wherein the training feature sequence corresponding to each time information comprises: at least one training payment feature, and each training payment feature corresponds to a feature type;
[0073] The training payment feature set includes: a training feature sequence corresponding to the multiple time information.
[0074] Optionally, the prediction module is further used to obtain the fifth predicted payment feature according to the training payment feature set, the fourth predicted payment feature, the local feature attention weights of the various features, and the global time attention weight of the feature sequence corresponding to each time information;
[0075] The parameters of the recurrent neural network model also include: the global temporal attention weight.
[0076] Optionally, the payment prediction model includes: a second decision tree model;
[0077] The prediction module is further used to obtain a sixth predicted payment feature of the target game using the second decision tree model according to the predicted payment feature output by the optimized first decision tree and the training payment feature set;
[0078] The optimization module is further used to optimize the parameters of the second decision tree model according to the verification payment feature set and the sixth prediction payment feature;
[0079] The prediction module is also used to obtain the fifth predicted payment feature using the recurrent neural network model based on the predicted payment feature output by the optimized second decision tree and the training payment feature set.
[0080] Optionally, the device further comprises: an acquisition module and a division module, wherein:
[0081] The acquisition module is used to acquire a sample payment feature set of the target game;
[0082] The division module is used to divide the sample payment feature set into multiple feature sets according to a preset division rule, and the multiple feature sets include: the training payment feature set and the verification payment feature set.
[0083] Optionally, the division module is further used to divide the sample payment feature information set into the multiple feature sets according to the time information of the payment features in the sample payment feature set and a preset division ratio.
[0084] Optionally, the plurality of feature sets further include: a test payment feature set;
[0085] The prediction module is further used to obtain prediction features of the first decision tree using the optimized first decision tree according to the test payment feature set;
[0086] The prediction module is further used to obtain the prediction features of the recurrent neural network using the optimized recurrent neural network model according to the prediction features of the first decision tree and the test payment feature set;
[0087] The calculation module is also used to perform weighted calculation according to the prediction features of the first decision tree and the prediction features of the recurrent neural network to obtain the payment prediction results of the target game during the test process, and the payment prediction results during the test process include: the prediction features after the weighted calculation.
[0088] In the fifth aspect, another embodiment of the present application provides a game payment prediction device, comprising: a processor, a storage medium and a bus, the storage medium storing machine-readable instructions executable by the processor, the processor and the storage medium communicating via the bus, the processor executing the machine-readable instructions to perform the steps of the method as described in any one of the first or second aspects above.
[0089] In a sixth aspect, another embodiment of the present application provides a storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in any one of the first aspect or the second aspect are executed.
[0090] The beneficial effects of the present application are: using the game payment prediction method provided by the present application, the payment feature set of the target game is predicted through the first decision tree model, and the first predicted payment feature and the payment feature set are input into the recurrent neural network model together to obtain the second predicted payment feature, and then weighted calculation is performed based on the first predicted payment feature and the second predicted payment feature to obtain the payment prediction result of the target game. Since the present application uses two models for prediction and performs weighted calculation on the results to obtain the final prediction result, and the recurrent neural network model obtains the second predicted payment feature based on the payment feature set, the first predicted payment feature and the local feature attention weights of various features, the model pays more attention to the features that have a greater impact on the output variable, so compared with traditional technologies, the results of the prediction model are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0091] 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.
[0092] Figure 1 A flowchart of a game payment prediction method provided in one embodiment of the present application;
[0093] Figure 2 A flowchart of a game payment prediction method provided by another embodiment of the present application;
[0094] Figure 3 A flowchart of a game payment prediction method provided by another embodiment of the present application;
[0095] Figure 4 A flowchart of a game payment prediction method provided by another embodiment of the present application;
[0096] Figure 5 A flowchart of a game payment prediction method provided by another embodiment of the present application;
[0097] Figure 6 A flowchart of a method for training a payment prediction model provided in one embodiment of the present application;
[0098] Figure 7 A flowchart of a method for training a payment prediction model provided in another embodiment of the present application;
[0099] Figure 8 A flowchart of a method for training a payment prediction model provided in another embodiment of the present application;
[0100] Fig. 9 A flowchart of a method for training a payment prediction model provided in another embodiment of the present application;
[0101] Fig.10 A schematic diagram of the structure of a game payment prediction device provided in one embodiment of the present application;
[0102] Fig.11 A schematic diagram of the structure of a game payment prediction device provided by another embodiment of the present application;
[0103] Fig.12 A schematic diagram of the structure of a training device for a payment prediction model provided in one embodiment of the present application;
[0104] Fig.13 A schematic diagram of the structure of a payment prediction model training device provided in another embodiment of the present application;
[0105] Fig.14 A schematic diagram of the structure of a payment prediction model training device provided in another embodiment of the present application;
[0106] Fig.15 A schematic diagram of the structure of a game payment prediction device provided in one embodiment of the present application. DETAILED DESCRIPTION
[0107] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.
[0108] Figure 1 A flowchart of a game payment prediction method provided in an embodiment of the present application is applied to a payment prediction model, the payment prediction model including: a first decision tree model and a recurrent neural network model. The game payment prediction method can be implemented by a game payment prediction device having the payment prediction model, and the game payment prediction device can be a payment prediction terminal or a payment prediction server.
[0109] like Figure 1 As shown, the method may include:
[0110] S101: According to the payment feature set of the target game, a first decision tree model is used to obtain a first predicted payment feature of the target game.
[0111] Optionally, the target game may be a game corresponding to the payment prediction model. In one embodiment of the present application, the first decision tree model may be Lightgbm or Xgboost. The specific selection of the first decision tree model may be determined according to user needs, and the present application does not impose any limitation thereto.
[0112] Optionally, in one embodiment of the present application, before using the first decision tree model for prediction, it is also necessary to perform data conversion operations on the payment feature set of the target game, including: taking a logarithm (log) operation, a normalization operation, and a preprocessing operation of converting the time series into a supervised problem.
[0113] Among them, the data with large numerical differences in the feature set will have a greater impact on the model. In order to eliminate the impact of large data fluctuations on the model, the logarithm of these data with large numerical differences is taken, and the logarithm will not change the correlation between these data. At the same time, it can make the values in the feature set more stable and reduce the abnormal sensitivity of different differences. However, for some data with small differences (such as churn rate, where the churn rate is a value between 0 and 1), it is not necessary to take the logarithm. The normalization operation is to improve the accuracy of the model and accelerate the convergence of the model. At the end of the preprocessing, the corresponding daily payment data is shifted upward by one time unit through the shift function operation (operation on moving data) of the data processing library pandas in Python, and the time series is converted into a supervised machine learning training problem (that is, the daily payment data corresponds to the payment data of the next day).
[0114] S102: According to the payment feature set and the first predicted payment feature, a recurrent neural network model is used to obtain a second predicted payment feature of the target game.
[0115] Among them, the preset recurrent neural network model is used to determine the local feature attention weights of various features in the payment feature set and the first predicted payment feature, and obtain the second predicted payment feature based on the payment feature set, the first predicted payment feature and the local feature attention weights of various features.
[0116] Optionally, in one embodiment of the present application, the recurrent neural network model can be, for example, a long short-term memory neural network (LSTM) model. The following embodiments are also described using the LSTM model as an example of the recurrent neural network model, but the selection of the specific recurrent neural network model can be designed according to user needs and is not limited to the model given in the above embodiment. The present application does not impose any restrictions on this.
[0117] Among them, the local feature attention weight is to guide the recurrent neural network model to pay more attention to the features that have a greater impact on the output variable, that is, the features that have a greater impact on the output variable should account for a higher proportion in the prediction process. The weights of different types of features may be the same or different. Such a setting will make the second predicted payment feature more accurate.
[0118] S103: Perform weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain a payment prediction result of the target game.
[0119] The payment prediction result includes: the predicted payment features obtained by weighted calculation.
[0120] Optionally, in one embodiment of the present application, the predicted payment feature is given according to the payment value per day.
[0121] It should be noted that the sum of the prediction weights of the first decision tree model and the recurrent neural network model is 1. For example: assuming that the prediction weight α of the first decision tree model is 0.4, the predicted paid income P1 is 1,000,000, the prediction weight γ of the recurrent neural network model is 0.6, and the paid income P2 is 1,080,000. The final prediction result of the model is P.
[0122] P=α*P1+γ*P2=0.4*1000000+0.6*1080000=1048000
[0123] The game payment prediction method provided by the present application is adopted to predict the payment of the payment feature set of the target game through the first decision tree model, and the first predicted payment feature and the payment feature set are input into the recurrent neural network model together to obtain the second predicted payment feature, and then the first predicted payment feature and the second predicted payment feature are weightedly calculated to obtain the payment prediction result of the target game. Since the present application adopts two models for prediction and performs weighted calculation on the results to obtain the final prediction result, and the improved recurrent neural network model obtains the second predicted payment feature based on the payment feature set, the first predicted payment feature and the local feature attention weights of various features, the model pays more attention to the features that have a greater impact on the output variable, so compared with traditional technologies, the results of the prediction model are more accurate.
[0124] Optionally, the payment feature set includes: at least one type of payment feature; at least one type of payment feature includes at least one of the following features: historical payment features, expected payment features in the future, and trend features corresponding to a preset prediction time step.
[0125] Among them, the historical payment characteristics may include, for example: basic operating indicators, payment structure, supply and demand, and social influence of the target game history for each day. For example, it may include: active number DAU, number of paying people, daily new number (Day New User, DNU), online time, average revenue per user (average revenue per user, ARPU), etc. In one embodiment of the present application, all players in the payment structure are divided into multiple levels according to the amount of historical payment and operating indicators, wherein the historical payment characteristics for each level may include: the number of logins and proportions of each VIP (very important person, VIP) level, historical recharge and proportion, statistical period recharge and proportion, number of people who have not logged in within a preset time period and proportion, historical number of people and churn rate and at least one feature. In the supply and demand influence, the main consideration is that when the payment trigger condition is "ingot", "silver", "gold coins" or "silver bills" and other virtual currency data in the game is insufficient, if the user's early consumption is too much, it may promote subsequent payment, so the user's consumption of different types of virtual currency and the distribution of consumption will also be added to the historical payment characteristics. Finally, we analyze the impact of social interaction on game payment, especially in Massive Multiplayer Online (MMO) mobile games. The economic system also has the phenomenon of birds of a feather flock together. Take guilds as an example, especially "guilds" with a high degree of concentration of big R players (key users who recharge a large amount of money in the game and recharge steadily). The number of active guild members and the number of people participating in guild activities every day are also characteristics of payment.
[0126] The expected payment features in the future may include: adding expected payment features in the future that have a significant and foreseeable impact on payment based on the periodic characteristics of the game, such as payment features corresponding to expected game events such as new servers, new activities, new versions, new copies, and holidays. The expected payment features in the future are discrete payment features that have a significant impact on payment, and often lead to a significant increase in game payment data for a period of time.
[0127] Among them, discrete features are relative to continuous features. Continuous features can take any value in a certain range, while discrete features can only be integers. For example: if a new server is opened on the same day, the discrete feature corresponding to the new server opening on the same day is 1; if no new server is opened on the same day, the discrete feature corresponding to the new server opening on the same day is 0. 0 and 1 are used to represent features such as whether a new server is opened on the same day.
[0128] The trend features corresponding to the preset prediction time step extract the statistical features of the translation time series, cumulative time series, etc. of the historical impact on the future with time trends according to the sliding time window of days, weeks, months, and multiple steps. For example: if the time window size is [7, 14, 30] and the object is the number of daily active users (DAU), the maximum, minimum, average, median, standard deviation, etc. of DAU in the last week, half a month, and last month can be extracted.
[0129] The local feature attention weights of various types of features include: the local feature attention weights of at least one type of payment feature, and the local feature attention weights of the first predicted payment feature.
[0130] That is, a local feature-level attention mechanism is added to the first predicted payment feature in each time step, the historical payment feature, the expected payment feature in the future, and the trend feature corresponding to the preset prediction time step, so as to construct a complete recurrent neural network model with a hybrid attention mechanism. Among them, the addition of discrete features enables the model to not only extract the impact of long-term game operation status on payment, but also add short-term activities, such as: factors that increase the price by opening new servers, so that the model can simultaneously capture the trend changes of game payment.
[0131] On the basis of the above method, an embodiment of the present application also provides an implementation example of a game payment prediction method, which can be explained by an implementation example of obtaining a payment feature set in the game payment prediction method. Figure 2 A flowchart of a game payment prediction method provided by another embodiment of the present application is shown in FIG. Figure 2 As shown, before S101, the method may further include:
[0132] S104: Sort the payment features in the payment feature set according to time to obtain feature sequences corresponding to multiple time information.
[0133] The feature sequence corresponding to each time information includes: at least one payment feature, and each payment feature corresponds to a feature type.
[0134] S105: The recurrent neural network model is used to obtain a second predicted payment feature based on the payment feature set, the first predicted payment feature, the local feature attention weights of each type of feature, and the global time attention weights of the feature sequence corresponding to each time information.
[0135] The method provided in the above embodiment is used to predict payment. Since the data in the recurrent neural network model includes a payment feature set and a first predicted payment feature, wherein the first predicted payment feature is predicted by the first decision tree model, and the recurrent neural network model also includes the local feature attention weights of various features, as well as the global time attention weights, the second predicted payment feature is calculated based on different weights. When the recurrent neural network model analyzes and processes the data, it will focus on the data that has a greater impact on the results, thereby making the prediction results more accurate. The payment feature set includes: a feature sequence corresponding to multiple time information.
[0136] Figure 3 FIG. 1 is a flow chart of a game payment prediction method provided in another embodiment of the present application. In another embodiment of the present application, the payment prediction model may further include: a second decision tree model. Figure 3 As shown, S102 may include:
[0137] S106: According to the payment feature set and the first predicted payment feature, a second decision tree model is used to obtain a third predicted payment feature of the target game.
[0138] Optionally, the first decision tree model and the second decision tree model can be decision tree models of the same type or different types. The selection and matching of specific decision tree models can be determined according to user needs, and this application does not impose any restrictions here.
[0139] Among them, since the payment features of the second decision tree not only include the original payment feature set, but also the first predicted payment features predicted by the first tree model, this is also an idea of boosting integration, which can be used to reduce learning bias in supervised learning.
[0140] In one embodiment of the present application, the first decision tree model and the second decision tree model are different types of decision tree models, wherein the first decision tree model is Lightgbm, and the second decision tree model is Xgboost.
[0141] S107: According to the payment feature set and the third predicted payment feature, a recurrent neural network model is used to obtain a second predicted payment feature.
[0142] In one embodiment of the present application, since the first two models are both tree models and the prediction results are not much different, they do not need to be input into the recurrent neural network model at the same time. That is, the recurrent neural network model contains the third predicted payment feature predicted by the second decision tree model, so there is no need to repeatedly add the first predicted payment feature predicted by the first decision tree model, and the second decision tree model has already learned the results of the first decision tree model, so there is no need to repeatedly add it. However, the specific application form can be designed according to user needs and is not limited to the above embodiments.
[0143] Correspondingly, S103 may include:
[0144] S108: Perform weighted calculation on the first predicted payment feature, the second predicted payment feature, and the third predicted payment feature to obtain a payment prediction result of the target game.
[0145] At this time, since the third predicted payment feature predicted by the second decision tree model is added in the calculation process of the payment prediction result, and each predicted payment feature is weighted according to the preset weight, the payment prediction result obtained will be more accurate. It should be noted that the sum of the prediction weight of the first decision tree model, the prediction weight of the second decision tree model and the prediction weight of the recurrent neural network model is 1.
[0146] Figure 4 A flowchart of a method for predicting game payment is provided in another embodiment of the present application. Figure 4 As shown, first, the historical payment features in the payment feature set, the expected payment features in the future time, and the trend features corresponding to the preset prediction time step are input into the first decision tree model respectively, and then the first predicted payment features predicted by the first decision tree model and the original payment feature set are input into the second decision tree model together, and then the third predicted payment features predicted by the second decision tree model and the original payment features are input into the time-recurrent neural network model together to obtain the second predicted payment features; finally, the first predicted payment features, the second predicted payment features, and the third predicted payment features are weightedly calculated to obtain the final payment prediction result of the target game.
[0147] Figure 5 A flowchart of a game payment prediction method provided by another embodiment of the present application, wherein the second predicted payment feature includes: a predicted payment feature corresponding to at least one predicted time step; Figure 5 As shown, S103 may include:
[0148] S109: Perform weighted calculation on the first predicted payment feature and the predicted payment feature corresponding to each predicted time step to obtain a payment prediction result corresponding to each predicted time step.
[0149] In one embodiment of the present application, when the second predicted payment feature includes a predicted payment feature corresponding to at least one predicted time step, the recurrent neural network model takes the sequence of the recent n weeks as a whole to construct a recurrent neural network model with multiple time steps; wherein each time series sample corresponds to a feature sequence for each day.
[0150] At this time, the input of the recurrent neural network is (batch, timestamps, features_dims), where batch is the time series sample size for each training, timestamps is the step size in each time series sample, and features_dims is the feature dimension size contained in each step.
[0151] At this time, since the recurrent neural network model uses a sequence-to-sequence implementation method, the corresponding input multi-step needs to pass through the encoder to pay attention to different types of features. For example: In one embodiment of the present application, for each time step input feature F = (f1, f2, ..., fT), it is divided into 4 parts including historical payment features F1 = (Fa1, Fa2, ..., FaT), expected payment features in the future time F2 = (Fb1, Fb2, ..., FbT), preset prediction time step corresponding trend features F3 = (Fc1, Fc2, ..., FcT) and predicted payment features F4 = (Fd1, Fd2, ..., FdT).
[0152] The encoder part analyzes that the impact of different prediction dates F1, F2, F3 and F4 on payment is significantly different. For example, if a new service will be launched tomorrow, the weight of the expected payment feature in the future should be significantly higher than other features. That is, local feature attention gives different attention weights to each type of feature, making the prediction result more accurate.
[0153] The payment prediction method provided by the present application includes a first tree model and a second tree model. Since the payment features of the second decision tree not only have the original payment feature set, but also the first predicted payment features predicted by the first tree model, the learning bias in supervised learning is reduced; at this time, the recurrent neural network model adds the third predicted payment features predicted by the second decision tree model in the calculation process of the payment prediction result, and performs weighted calculation on each predicted payment feature according to the preset weight, so the payment prediction result obtained will be more accurate, making the prediction model more generalized. It makes up for the shortcomings of the tree model that it is insensitive to time trends and weakens long-term storage of information. At the same time, a hybrid attention mechanism of global time attention and local feature attention is introduced into the recurrent neural network model, which strengthens the capture of factors strongly related to payment by the recurrent neural network model. The global time attention mechanism can guide the model to pay more attention to the sequence that has a greater impact on the output sequence, and the addition of a local feature attention mechanism in each time series can guide the model to pay more attention to the features that have a greater impact on the output variable, so that the results of the prediction model can be more accurate. In addition, the recurrent neural network model also includes the local feature attention weights of various features, as well as the global time attention weights, so the second predicted payment feature is calculated based on different weights, which strengthens the recurrent neural network model's capture of factors that are strongly correlated with payment. Among them, adding a global time attention mechanism to each time series will make the model pay more attention to the sequence that has a greater impact on the output sequence, and adding a local feature attention mechanism to each time series will make the model pay more attention to the features that have a greater impact on the output variable, which can make the prediction model more accurate.
[0154] To achieve the above Figure 1 - Figure 5 In any of the above-mentioned game payment prediction methods, the present application embodiment also provides a payment prediction model training method to train a payment prediction model that can execute the above-mentioned game payment prediction method. An example is used to explain the above-mentioned payment prediction model.
[0155] Figure 6 A flow chart of a payment prediction model training method provided in one embodiment of the present application, wherein the payment prediction model includes: a first decision tree model and a recurrent neural network model. The payment prediction model training method can be implemented by a game payment prediction training device having the payment prediction model, and the game payment prediction training device can be a payment prediction training terminal or a payment prediction server, such as Figure 6 As shown, the model training method includes:
[0156] S201: According to the training payment feature set of the target game, the first decision tree model is used to obtain the fourth predicted payment feature of the target game.
[0157] In one embodiment of the present application, the data in the training payment feature set is data that has undergone a data conversion operation. The specific data conversion operation method is the same as that described above. Figure 1 - Figure 4 The method given is the same and will not be repeated here.
[0158] S202: Optimizing the parameters of the first decision tree according to the verified payment feature set and the fourth predicted payment feature of the target game.
[0159] Among them, the process of optimizing the parameters of the first decision tree is to adjust the parameters using the GridSearch method to obtain the first decision tree with the best performance, and then predict the training payment feature set again based on the optimized first decision tree to obtain the output predicted payment features.
[0160] S203: Based on the predicted payment feature output by the optimized first decision tree and the training payment feature set, a recurrent neural network model is used to obtain the fifth predicted payment feature of the target game.
[0161] S204: Optimizing parameters of the recurrent neural network model according to the verification payment feature set and the fifth prediction payment feature.
[0162] The parameters of the preset recurrent neural network model include: input_size, the dimension of the input feature, the number of layers of the LSTM stack, the dimension of the hidden layer, etc. This application does not impose any restrictions here.
[0163] S205: Perform weighted calculation on the predicted payment features output by the optimized first decision tree and the predicted payment features output by the optimized recurrent neural network model.
[0164] Among them, the result of weighted calculation is the final payment prediction result.
[0165] The training method of the payment prediction model provided in the present application is adopted. Since the first decision tree model is obtained by continuously optimizing the verification payment feature set and the training payment feature set of the target game, and the predicted payment features output by the optimized first decision tree model are added to the recurrent neural network model, and the parameters of the recurrent neural network model are optimized according to the verification payment feature set and the fifth predicted payment features to obtain the optimized recurrent neural network model, then, the predicted payment features output by the optimized first decision tree and the predicted payment features output by the optimized recurrent neural network model are weightedly calculated to obtain the final payment prediction result. Such a training method will enable the first decision tree model and the recurrent neural network model to automatically optimize the parameters, so that the final models are all optimized models, which are convenient for use in subsequent predictions.
[0166] Optionally, the training payment feature set includes: at least one type of training payment feature; the local feature attention weights of each type of feature include: the local feature attention weights of at least one type of training payment feature, and the local feature attention weights of the fourth prediction payment feature.
[0167] Figure 7 A flowchart of a method for training a payment prediction model provided in another embodiment of the present application is shown in FIG. Figure 7 As shown, before S201, the method further includes:
[0168] S206: Sort the training payment features in the training payment feature set according to time to obtain a plurality of training feature sequences corresponding to time information.
[0169] The training feature sequence corresponding to each time information includes: at least one training payment feature, each training payment feature corresponds to a feature type. The training payment feature set includes: a plurality of training feature sequences corresponding to time information.
[0170] Correspondingly, S203 may include:
[0171] S207: The recurrent neural network model obtains the fifth predicted payment feature according to the training payment feature set, the fourth predicted payment feature, the local feature attention weights of various features, and the global time attention weights of the feature sequence corresponding to each time information.
[0172] The parameters of the recurrent neural network model also include: global temporal attention weight.
[0173] Among them, the global time attention weight can guide the model to pay more attention to data that has a greater impact on the results. For example, if you need to predict the payment value of this year's Double Eleven, the payment value on Double Eleven last year has a greater impact on the result than other data, while the data of the two days closest to the current prediction date are not very referenceable. Therefore, different attention weights need to be set for data corresponding to different times.
[0174] Figure 8 A flow chart of a method for training a payment prediction model provided in another embodiment of the present application, wherein the payment prediction model includes: a second decision tree model; Figure 8 As shown, S203 includes:
[0175] S208: Based on the predicted payment feature output by the optimized first decision tree and the training payment feature set, the second decision tree model is used to obtain the sixth predicted payment feature of the target game.
[0176] S209: Optimizing parameters of the second decision tree model according to the verified payment feature set and the sixth predicted payment feature.
[0177] S210: According to the predicted payment feature and the training payment feature set output by the optimized second decision tree, a recurrent neural network model is used to obtain a fifth predicted payment feature.
[0178] Optionally, for example: assuming that the first decision tree model prediction weight α is 0.2, the predicted paid income P1 is 1000000, the second decision tree model prediction weight β is 0.2, the predicted paid income P2 is 1100000, the recurrent neural network model prediction weight γ is 0.6, the predicted paid income P3 is 1080000, and the final model prediction result is P.
[0179] P=α*P_1+β*P_2+γ*P_3=0.2*1000000+0.2*1100000+0.6*1080000=1068000
[0180] However, the weight setting of the prediction results of each specific model can be designed according to user needs and is not limited to the above embodiment.
[0181] Fig. 9 A flowchart of a method for training a payment prediction model provided in another embodiment of the present application is shown in FIG. Fig. 9 As shown, before S201, the method further includes:
[0182] S211: Obtain a sample payment feature set of the target game.
[0183] S212: Divide the sample payment feature set into multiple feature sets according to a preset division rule.
[0184] The multiple feature sets include: a training paid feature set and a verification paid feature set.
[0185] Optionally, in one embodiment of the present application, the sample payment feature information set may be divided into multiple feature sets according to the time information of the payment features in the sample payment feature set and a preset division ratio.
[0186] Optionally, in one embodiment of the present application, the target game's to-be-trained sample set includes: a training payment feature set and a verification payment feature set; after obtaining the target game's to-be-trained sample set, the training sample set needs to be treated according to the above Figure 1The method provided performs data conversion operation. Then, the sample set to be trained after data conversion is sorted in chronological order, and the sample set to be trained after data conversion is divided into a training paid feature set and a verification paid feature set according to a preset ratio; wherein the preset ratio can be adjusted according to user needs. In one embodiment of the present application, the preset ratio is 7:3, that is, the sample set to be trained after data conversion and sorting is divided into a training paid feature set in chronological order from front to back, and the first 70% is divided into a verification paid feature set, and the last 30% is divided into a verification paid feature set; the training paid feature set and the verification paid feature set are mainly used to calculate the loss function value, so as to optimize the fitting of the parameters of the first decision tree, until the loss function value obtained based on the first decision tree after the optimized fitting parameters converges.
[0187] Among them, in one embodiment of the present application, the predicted payment feature is obtained after the optimized first decision tree model processes the training payment feature set, or it can be obtained after the optimized first decision tree model processes the training sample set of the target game. It can be specifically designed according to user needs and is not limited to the above two embodiments.
[0188] Optionally, in one embodiment of the present application, the plurality of feature sets further include: a test payment feature set.
[0189] After obtaining the target game's training sample set, you need to treat the training sample set according to the above Figure 1 The method provided performs a data conversion operation. The sample set to be trained after data conversion is then sorted in chronological order, and the sample set to be trained after data conversion is divided into a training paid feature set, a verification paid feature set, and a test paid feature set according to a preset ratio; wherein the preset ratio can be adjusted according to user needs. In one embodiment of the present application, the preset ratio is 6:2:2, that is, the sample set to be trained after data conversion and sorting is divided into a training paid feature set in chronological order from front to back, the first 60% is divided into a verification paid feature set, the middle 20% is divided into a verification paid feature set, and the last 20% is divided into a test paid feature set.
[0190] During the training process, the training payment feature set and the verification payment feature set are mainly used to calculate the loss function value, so as to optimize the fitting of the parameters of the first decision tree until the loss function value obtained based on the first decision tree after the optimized fitting parameters converges, and then the optimized first decision tree is used according to the test payment feature set to obtain the prediction features of the first decision tree. Subsequently, according to the prediction features of the first decision tree and the test payment feature set, the optimized recurrent neural network model is used to obtain the prediction features of the recurrent neural network. Finally, a weighted calculation is performed based on the prediction features of the first decision tree and the prediction features of the recurrent neural network to obtain the payment prediction results of the target game during the test process. The payment prediction results during the test process include: the prediction features after weighted calculation.
[0191] The training method of the payment prediction model provided in this application is adopted. Since each model is obtained after self-optimization, and the prediction data output by the optimized model will be used as new sample data and input into the next model together with the original sample payment feature set, the prediction model has more generalization performance. It makes up for the shortcomings of the tree model that is insensitive to time trends and weakened by long-term storage of information. At the same time, a hybrid attention mechanism of global time attention and local feature attention is introduced into the recurrent neural network model, which strengthens the recurrent neural network model's capture of factors strongly related to payment. The global time attention mechanism can guide the model to pay more attention to sequences that have a greater impact on the output sequence. The addition of a local feature attention mechanism in each time series can guide the model to pay more attention to features that have a greater impact on the output variable, which can make the results of the prediction model more accurate.
[0192] Fig.10 A schematic diagram of the structure of a game payment prediction device provided in an embodiment of the present application is applied to a payment prediction model, such as Fig.10 As shown, the payment prediction model includes: a first decision tree model and a recurrent neural network model; the device includes: a prediction module 301 and a calculation module 302, wherein:
[0193] The prediction module 301 is used to obtain the first predicted payment feature of the target game by using the first decision tree model according to the payment feature set of the target game.
[0194] The prediction module 301 is used to obtain the second predicted payment feature of the target game based on the payment feature set and the first predicted payment feature by using a recurrent neural network model; wherein the preset recurrent neural network model is used to determine the local feature attention weights of various features in the payment feature set and the first predicted payment feature, and obtain the second predicted payment feature based on the payment feature set, the first predicted payment feature and the local feature attention weights of various features.
[0195] The calculation module 302 is used to perform weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain a payment prediction result of the target game; wherein the payment prediction result includes: the predicted payment feature obtained by weighted calculation.
[0196] Fig.11 A schematic diagram of the structure of a game payment prediction device provided in another embodiment of the present application is applied to a payment prediction model, such as Fig.11 As shown, the device also includes: a sorting module 303, which is used to sort the payment features in the payment feature set according to time to obtain feature sequences corresponding to multiple time information, and the feature sequence corresponding to each time information includes: at least one payment feature, and each payment feature corresponds to a feature type.
[0197] The paid feature set includes: feature sequences corresponding to multiple time information.
[0198] Optionally, the prediction module 301 is also used to obtain a second predicted payment feature based on the payment feature set, the first predicted payment feature, the local feature attention weights of various features, and the global time attention weights of the feature sequence corresponding to each time information.
[0199] Optionally, the prediction module 301 is further configured to obtain a third predicted payment feature of the target game using a second decision tree model according to the payment feature set and the first predicted payment feature;
[0200] The prediction module 301 is further used to obtain a second predicted payment feature by using a recurrent neural network model according to the payment feature set and the third predicted payment feature;
[0201] The calculation module 302 is further used to perform weighted calculation on the first predicted payment feature, the second predicted payment feature and the third predicted payment feature to obtain a payment prediction result of the target game.
[0202] Optionally, the second predicted payment feature includes: a predicted payment feature corresponding to at least one predicted time step; the calculation module 302 is also used to perform weighted calculation on the first predicted payment feature and the predicted payment feature corresponding to each predicted time step to obtain a payment prediction result corresponding to each predicted time step.
[0203] The above-mentioned device is used to perform the above-mentioned Figure 1 - Figure 4 The implementation principle and technical effect of the method provided in the embodiment are similar and will not be described in detail here.
[0204] Fig.12 A schematic diagram of a payment prediction model training device provided in an embodiment of the present application is shown in FIG. Fig.12As shown, the payment prediction model includes: a first decision tree model and a recurrent neural network model, and the device includes: a prediction module 401, an optimization module 402 and a calculation module 403, wherein:
[0205] The prediction module 401 is used to obtain the fourth predicted payment feature of the target game by using the first decision tree model according to the training payment feature set of the target game.
[0206] The optimization module 402 is used to optimize the parameters of the first decision tree according to the verification payment feature set and the fourth predicted payment feature of the target game.
[0207] The prediction module 401 is also used to obtain the fifth predicted payment feature of the target game using a recurrent neural network model based on the predicted payment feature output by the optimized first decision tree and the training payment feature set.
[0208] The optimization module 402 is further used to optimize the parameters of the recurrent neural network model according to the verification payment feature set and the fifth predicted payment feature.
[0209] The calculation module 403 is used to perform weighted calculation on the predicted payment features output by the optimized first decision tree and the predicted payment features output by the optimized recurrent neural network model.
[0210] Fig.13 A schematic diagram of a payment prediction model training device provided in another embodiment of the present application is shown in FIG. Fig.13 As shown, the device also includes a sorting module 404, which is used to sort each training payment feature in the training payment feature set according to time to obtain a training feature sequence corresponding to multiple time information, and the training feature sequence corresponding to each time information includes: at least one training payment feature, and each training payment feature corresponds to a feature type.
[0211] The training payment feature set includes: a training feature sequence corresponding to multiple time information.
[0212] Optionally, the prediction module 401 is also used to obtain a fifth predicted payment feature based on the training payment feature set, the first predicted payment feature, the local feature attention weights of each type of feature, and the global time attention weight of the feature sequence corresponding to each time information.
[0213] The parameters of the recurrent neural network model also include: global temporal attention weight.
[0214] Optionally, the payment prediction model includes: a second decision tree model; a prediction module 401, which is also used to obtain the sixth predicted payment feature of the target game using the second decision tree model based on the predicted payment feature output by the optimized first decision tree and the training payment feature set.
[0215] The optimization module 402 is further used to optimize the parameters of the second decision tree model according to the verification payment feature set and the sixth predicted payment feature.
[0216] The prediction module 401 is further used to obtain the fifth predicted payment feature by adopting a recurrent neural network model according to the training payment feature set and the sixth predicted payment feature.
[0217] Fig.14 A schematic diagram of a payment prediction model training device provided in another embodiment of the present application is shown in FIG. Fig.14 As shown, the device further includes: an acquisition module 405 and a division module 406, wherein:
[0218] The acquisition module 405 is used to acquire a sample payment feature set of the target game.
[0219] The division module 406 is used to divide the sample payment feature set into multiple feature sets according to a preset division rule. The multiple feature sets include: a training payment feature set and a verification payment feature set.
[0220] Optionally, the division module 406 is further used to divide the sample payment feature information set into multiple feature sets according to the time information of the payment feature in the sample payment feature set and a preset division ratio.
[0221] Optionally, the multiple feature sets further include: a test payment feature set.
[0222] The prediction module 401 is further used to obtain prediction features of the first decision tree using the optimized first decision tree according to the test payment feature set.
[0223] The prediction module 401 is also used to obtain the prediction features of the recurrent neural network using the optimized recurrent neural network model according to the prediction features of the first decision tree and the test payment feature set.
[0224] The calculation module 403 is also used to perform weighted calculation based on the prediction features of the first decision tree and the prediction features of the recurrent neural network to obtain the payment prediction results of the target game during the test process. The payment prediction results during the test process include: the prediction features after weighted calculation.
[0225] The above-mentioned device is used to perform the above-mentioned Figure 5 - Figure 8 The implementation principle and technical effect of the method provided in the embodiment are similar and will not be described in detail here.
[0226] The above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASIC), or one or more digital singnal processors (DSP), or one or more field programmable gate arrays (FPGA). For another example, when a module is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).
[0227] Fig.15 A structural diagram of a game payment prediction device provided for one embodiment of the present application. The game payment prediction device can be integrated into a game payment prediction terminal, or a chip of a game payment prediction terminal, a game payment prediction server, or a chip of a game payment prediction server.
[0228] The game payment prediction device includes: a processor 501 , a storage medium 502 and a bus 503 .
[0229] The storage medium 502 stores machine-readable instructions executable by the processor 501. The processor 501 communicates with the storage medium 502 via a bus 503. The processor 501 executes the machine-readable instructions to perform the above Figure 1 - Fig. 9 The specific implementation methods and technical effects are similar and will not be described in detail here.
[0230] Optionally, the present application also provides a program product, such as a storage medium, on which a computer program is stored, including a program, which executes the above-mentioned Figure 1 - Fig. 9 An embodiment corresponding to any of the methods shown.
[0231] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as 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 interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.
[0232] 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.
[0233] 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. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional units.
[0234] The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (English: processor) to perform some 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 (English: Read-Only Memory, abbreviated: ROM), random access memory (English: Random Access Memory, abbreviated: RAM), disk or optical disk and other media that can store program codes.
Claims
1. A game payment prediction method, characterized in that: Applied to a payment prediction model, the payment prediction model includes: a first decision tree model and a preset recurrent neural network model; the method includes: According to the payment feature set of the target game, the first decision tree model is used to obtain a first predicted payment feature of the target game; According to the payment feature set and the first predicted payment feature, the recurrent neural network model is used to obtain a second predicted payment feature of the target game; wherein the preset recurrent neural network model is used to determine the local feature attention weights of various features in the payment feature set and the first predicted payment feature, and obtain the second predicted payment feature according to the payment feature set, the first predicted payment feature and the local feature attention weights of various features; The first predicted payment feature and the second predicted payment feature are weightedly calculated to obtain the payment prediction result of the target game; the payment feature set includes at least one of the following features: historical payment features, expected payment features in the future, and trend features; wherein the historical payment features include: basic operating indicators, payment structure, supply and demand, and social influence for each day in the history of the target game; wherein the supply and demand influence includes the consumption of different types of virtual currencies and the distribution of consumption; the expected payment features in the future include: adding expected payment features in the foreseeable future based on the cyclical characteristics of the game, the expected payment features in the future are payment features corresponding to expected game events, including: opening new servers, new activities, new versions, new copies, and holidays; the trend features include: extracting the translational time series and cumulative time series of the historical impact on the future with time trends based on sliding time windows of days, weeks, months, and multiple steps.
2. The method according to claim 1, characterized in that The payment feature set includes: at least one type of payment feature; The local feature attention weights of each type of feature include: the local feature attention weight of the at least one type of payment feature, and the local feature attention weight of the first predicted payment feature.
3. The method according to claim 1, characterized in that Before obtaining the first predicted payment feature of the target game by using the first decision tree model according to the payment feature set of the target game, the method further includes: Sorting the payment features in the payment feature set according to time to obtain a plurality of feature sequences corresponding to time information, wherein each feature sequence corresponding to time information includes: at least one payment feature, and each payment feature corresponds to a feature type; The payment feature set includes: feature sequences corresponding to the multiple time information.
4. The method according to claim 3, characterized in that The method further comprises: The recurrent neural network model is used to obtain the second predicted payment feature based on the payment feature set, the first predicted payment feature, the local feature attention weights of each type of feature, and the global time attention weights of the feature sequence corresponding to each time information.
5. The method according to claim 1, characterized in that The payment prediction model includes: a second decision tree model, wherein the recurrent neural network model is used according to the payment feature set and the first predicted payment feature, including: According to the payment feature set and the first predicted payment feature, the second decision tree model is used to obtain a third predicted payment feature of the target game; According to the payment feature set and the third predicted payment feature, the recurrent neural network model is used to obtain the second predicted payment feature; The step of performing weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain the payment prediction result of the target game includes: The first predicted payment feature, the second predicted payment feature and the third predicted payment feature are weightedly calculated to obtain a payment prediction result of the target game.
6. The method according to claim 1, characterized in that The second predicted payment feature includes: a predicted payment feature corresponding to at least one predicted time step; The step of performing weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain the payment prediction result of the target game includes: A weighted calculation is performed on the first predicted payment feature and the predicted payment feature corresponding to each predicted time step to obtain a payment prediction result corresponding to each predicted time step.
7. A method for training a payment prediction model, characterized in that: The payment prediction model includes: a first decision tree model and a preset recurrent neural network model, and the training method of the model includes: According to the training payment feature set of the target game, the first decision tree model is used to obtain the fourth predicted payment feature of the target game; the training payment feature set includes: at least one type of training payment feature, and the training payment feature set includes at least one of the following features: historical payment features, expected payment features in the future, and trend features; wherein the historical payment features include: basic operating indicators, payment structure, supply and demand, and social influence on each day of the target game history; wherein the supply and demand influence includes the consumption of different types of virtual currencies and the distribution of consumption; the expected payment features in the future include: adding expected payment features in the foreseeable future time according to the periodic characteristics of the game, and the expected payment features in the future time are payment features corresponding to expected game events, including: opening new servers, new activities, new versions, new copies, and holidays; the trend features include: extracting the translation time series and cumulative time series of the historical impact on the future with time trends according to the sliding time window of days, weeks, months, and multiple steps; Optimizing the parameters of the first decision tree according to the verified payment feature set of the target game and the fourth predicted payment feature; According to the predicted payment feature output by the optimized first decision tree and the training payment feature set, the recurrent neural network model is used to obtain a fifth predicted payment feature of the target game; Optimizing the parameters of the recurrent neural network model according to the verification payment feature set and the fifth prediction payment feature; A weighted calculation is performed on the predicted payment features output by the optimized first decision tree and the predicted payment features output by the optimized recurrent neural network model.
8. The method according to claim 7, characterized in that The local feature attention weights of various types of features include: the local feature attention weights of the at least one type of training payment feature, and the local feature attention weights of the fourth predicted payment feature.
9. The method according to claim 7, characterized in that Before obtaining the fourth predicted payment feature of the target game by using the first decision tree model according to the training payment feature set of the target game, the method further includes: Sort the training payment features in the training payment feature set according to time to obtain a plurality of training feature sequences corresponding to time information, wherein each training feature sequence corresponding to time information includes at least one training payment feature, and each training payment feature corresponds to a feature type; The training payment feature set includes: a training feature sequence corresponding to the multiple time information.
10. The method according to claim 9, characterized in that The method further comprises: The recurrent neural network model is used to obtain the fifth predicted payment feature according to the training payment feature set, the fourth predicted payment feature, the local feature attention weights of various features, and the global time attention weights of the feature sequence corresponding to each time information; The parameters of the recurrent neural network model also include: the global temporal attention weight.
11. The method according to claim 7, characterized in that The payment prediction model includes: a second decision tree model; the predicted payment feature outputted by the optimized first decision tree and the training payment feature set, using the recurrent neural network model to obtain the fifth predicted payment feature of the target game, including: According to the predicted payment feature output by the optimized first decision tree and the training payment feature set, the second decision tree model is used to obtain a sixth predicted payment feature of the target game; Optimizing the parameters of the second decision tree model according to the verification payment feature set and the sixth prediction payment feature; According to the predicted payment feature output by the optimized second decision tree and the training payment feature set, the recurrent neural network model is adopted to obtain the fifth predicted payment feature.
12. The method according to any one of claims 7 to 11, characterized in that: Before obtaining the fourth predicted payment feature of the target game by using the first decision tree model according to the training payment feature set of the target game, the method further includes: Obtaining a sample payment feature set of the target game; According to a preset division rule, the sample payment feature set is divided into a plurality of feature sets, and the plurality of feature sets include: the training payment feature set and the verification payment feature set.
13. The method according to claim 12, characterized in that According to the preset division rules, the sample payment feature set is divided into multiple feature sets, including: According to the time information of the payment features in the sample payment feature set and a preset division ratio, the sample payment feature set is divided into the multiple feature sets.
14. The method according to claim 12, characterized in that The plurality of feature sets also include: a test payment feature set; According to the test payment feature set, using the optimized first decision tree, obtaining prediction features of the first decision tree; According to the prediction features of the first decision tree and the test payment feature set, the optimized recurrent neural network model is used to obtain the prediction features of the recurrent neural network; A weighted calculation is performed based on the prediction features of the first decision tree and the prediction features of the recurrent neural network to obtain a payment prediction result of the target game during the test process, and the payment prediction result during the test process includes: the prediction features after the weighted calculation.
15. A game payment prediction device, applied to a payment prediction model, characterized in that: The payment prediction model includes: a first decision tree model and a preset recurrent neural network model; the device includes: a prediction module and a calculation module, wherein: The prediction module is used to obtain a first predicted payment feature of the target game using the first decision tree model according to the payment feature set of the target game; The prediction module is used to obtain the second predicted payment feature of the target game using the recurrent neural network model according to the payment feature set and the first predicted payment feature; wherein the preset recurrent neural network model is used to determine the local feature attention weights of various features in the payment feature set and the first predicted payment feature, and obtain the second predicted payment feature according to the payment feature set, the first predicted payment feature and the local feature attention weights of various features; The calculation module is used to perform weighted calculation on the first predicted payment feature and the second predicted payment feature to obtain the payment prediction result of the target game; the payment feature set includes at least one of the following features: historical payment features, expected payment features in the future, and trend features; wherein the historical payment features include: basic operating indicators, payment structure, supply and demand, and social influence for each day in the history of the target game; wherein the supply and demand influence includes the consumption of different types of virtual currencies and the distribution of consumption; the expected payment features in the future include: adding expected payment features in the foreseeable future based on the periodic characteristics of the game, and the expected payment features in the future are payment features corresponding to expected game events, including: opening new servers, new activities, new versions, new copies, and holidays; the trend features include: extracting the translation time series and cumulative time series of the historical impact on the future with time trends based on sliding time windows of days, weeks, months, and multiple steps.
16. A training device for a payment prediction model, characterized in that: The payment prediction model includes: a first decision tree model and a preset recurrent neural network model, and the device includes: a prediction module, an optimization module and a calculation module, wherein: The prediction module is used to obtain the fourth predicted payment feature of the target game by using the first decision tree model according to the training payment feature set of the target game; the training payment feature set includes: at least one type of training payment feature, and the training payment feature set includes at least one of the following features: historical payment feature, expected payment feature in the future time, and trend feature; wherein the historical payment feature includes: basic operating indicators, payment structure, supply and demand, and social influence of each day in the history of the target game; wherein the supply and demand influence includes the consumption of different types of virtual currencies and the distribution of consumption; the expected payment feature in the future time includes: adding the expected payment feature in the foreseeable future time according to the periodic characteristics of the game, and the expected payment feature in the future time is the payment feature corresponding to the expected game event, including: opening a new server, a new activity, a new version, a new copy, and holidays; the trend feature includes: extracting the translation time series and cumulative time series of the historical impact on the future with time trend according to the sliding time window of day, week, month, and multiple steps; The optimization module is used to optimize the parameters of the first decision tree according to the verification payment feature set of the target game and the fourth predicted payment feature; The prediction module is further used to obtain a fifth predicted payment feature of the target game using the recurrent neural network model according to the predicted payment feature output by the optimized first decision tree and the training payment feature set; The optimization module is further used to optimize the parameters of the recurrent neural network model according to the verification payment feature set and the fifth prediction payment feature; The calculation module is used to perform weighted calculation on the predicted payment features output by the optimized first decision tree and the predicted payment features output by the optimized recurrent neural network model.
17. A game payment prediction device, characterized in that: include: A processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, the processor and the storage medium communicate via the bus, and the processor executes the machine-readable instructions to execute the method described in any one of claims 1 to 14.
18. A storage medium having a computer program stored thereon, wherein the computer program executes the method according to any one of claims 1 to 14 when executed by a processor.
