Internet marketing decision system and method based on deep learning optimization

Abstract

The invention relates to the technical field of marketing decision making, and discloses an internet marketing decision making system and method based on deep learning optimization, and the internet marketing decision making method based on deep learning optimization comprises the following steps: carrying out the independent division of each game, and setting an independent total budget; first comprehensive information is collected; encoding and inputting into an allocation strategy model, and outputting an allocation decision; and decoding and integrating the distribution decision of each game to obtain a complete marketing decision. When the game is adjusted, second comprehensive information of the game is collected; encoding and inputting into an influence prediction model, and outputting an influence level; and judging whether the game needs to be independently separated from the marketing decision-making system or not according to the output influence prediction grade of the game. According to the method, the game is divided independently, a judgment method for the influence on currency transaction is added, when it is judged that the game can be greatly influenced, the game can be separated independently, and meanwhile execution of marketing decisions of other games is not influenced.

Description

Technical Field

[0001] The present invention relates to the technical field of marketing decision-making, and more specifically, to an Internet marketing decision-making system and method based on deep learning optimization. Background Art

[0002] An intermediary service company focuses on the field of online game virtual currency transactions, serving as an important link between game users and game operators. The company adopts a business model of purchasing virtual currency on a large scale and selling it to end users at competitive prices, earning profits through price differences. However, in the increasingly fierce market competition environment, the company urgently needs to optimize the marketing strategy decision-making system to achieve continuous improvement in sales performance and profit levels.

[0003] Traditional decision-making methods mainly rely on model analysis, and maximize profits by continuously adjusting the resource investment ratio of each game project. However, when game developers make major adjustments to the game's economic system, such as introducing new virtual currencies, changing the currency acquisition mechanism, or innovating consumption channels, it may significantly affect the game's life cycle curve, causing challenges to the applicability of the original decision-making model. Summary of the invention

[0004] The present invention provides an Internet marketing decision-making system and method based on deep learning optimization, which solves the technical problem in the related technology that when one or more game developers make major adjustments to the game economic system, the original model becomes unapplicable and the trading strategies of other games cannot be adjusted within a short period of time.

[0005] The present invention provides an Internet marketing decision-making method based on deep learning optimization, comprising the following steps: Step 100, divide each game into separate categories and set a separate total budget for each game; Step 200, collecting the first comprehensive information of each game at n consecutive time points before the current time point, the first comprehensive information including: sales data, player data and market data; The data is collected at fixed time intervals; Step 300, generating No. 1 plane data for each game based on No. 1 comprehensive information encoding of a total of n consecutive time points before and including the current time point of each game; Step 400, inputting the No. 1 plane data of each game into an allocation strategy model, the allocation strategy model comprising a first encoder, a second encoder and a first fully connected layer, wherein the first encoder is used to input the No. 1 plane data, output a first identification feature to the second encoder, the second encoder outputs a second identification feature to the first fully connected layer, and the first fully connected layer outputs an allocation decision for each game; Step 500, decoding and integrating the allocation decisions of each game to obtain a complete marketing decision.

[0006] In a preferred embodiment, the game that needs to be adjusted when executing the marketing decision is defined as Game A, and the influence of the adjustment of Game A on the marketing of virtual currency is determined by the following steps, including: Step 101, collecting the second comprehensive information of the game A at n consecutive time points before the current time point, the second comprehensive information including: adjustment information, player data, player feedback information, and social media information; Adjustment information includes: update logs, virtual currency adjustment information; Step 102, generating a second serial data stream based on the second comprehensive information encoding of n consecutive time points before and including the current time point, wherein the second serial data stream of a second data object comprises n second data units sorted in time, and the t-th second data unit represents the second comprehensive information collected at the t-th moment of the second data object; Step 103, inputting the second serial data stream into the impact prediction model, the impact prediction model comprising a third encoder and a second fully connected layer, wherein the third encoder is used to input the second serial data stream, output the third identification feature to the second fully connected layer, and the second fully connected layer outputs the impact level; Step 104: determine whether it is necessary to separate the game from the marketing decision system based on the outputted predicted impact level of the game.

[0007] In a preferred embodiment, the output of the impact prediction model is represented as a vector, the components of the vector respectively represent the probability of each impact level of game A at the next time point, and the impact level with the highest probability is taken as the output.

[0008] In a preferred embodiment, the second comprehensive information is input into the impact prediction model after feature engineering; the second comprehensive information includes text modality, and the Word Embedding algorithm is used as a feature engineering method.

[0009] In a preferred embodiment, the impact prediction model uses a multi-classification cross entropy loss function as the loss function of the model, and the calculation formula is as follows:

[0010] in represents the loss value, M represents the total number of classification categories, represents the total number of training samples, It is The symbol value of the training sample, if the The true classification of the training sample is the cth classification category and its value is 1, otherwise it is 0. It is the first factor that affects the prediction model. The probability that a training sample belongs to the cth classification category, Represents the logarithmic function with the natural constant e as the base.

[0011] In a preferred embodiment, the sales data includes sales revenue, sales cost, and sales volume; Player data includes the number of players, activity, and payment rate; Market data includes market size, growth rate, and competition.

[0012] In a preferred embodiment, the first plane data includes a first bin for storing data, a first bin indexes a first serial data stream to a first data object, and the first plane data also includes a first relationship list for storing association relationships between the first data objects indexed by the first bin, and the element value of the first relationship list in the i-th row and j-th column represents the quantized value of the association relationship between the first data objects indexed by the i-th and j-th first bins; The association relationships include high and low, corresponding to two quantitative values; Data objects include players, sales, markets, and games; The data objects with high correlation are defined as: games and players in the games, games and the sales of their virtual currency, games and the market for the sales of their virtual currency; A No. 1 serial data stream of a No. 1 data object includes n No. 1 data units sorted in time, and the t-th No. 1 data unit represents the No. 1 comprehensive information collected at the t-th moment of the No. 1 data object.

[0013] In a preferred embodiment, the output of the allocation strategy model represents a vector, and the components of the vector respectively represent the total proportion and the expenditure proportion. The total proportion represents the proportion of the total budget of the game to the total budget of all games, and the expenditure proportion represents the proportion of the actual investment in purchasing virtual currency for the game to the total budget of the game.

[0014] In a preferred embodiment, the allocation strategy model is trained using a reinforcement learning method, and the training steps of the allocation strategy model are as follows: Step 201, copy the allocation strategy model to generate the Actor network and the TargetActor network, use the feedforward neural network as the Critic network, copy the Critic network to obtain the TargetCritic network, and randomly initialize the parameters of the Actor network, the TargetActor network, the Critic network, and the TargetCritic network; Step 202, input The first comprehensive information collected Go to the Actor network to get the The allocation strategy for the execution of ; No. The allocation strategy for the execution of Add random noise to get The noise allocation strategy ; The first The allocation strategy for the execution of Input into the environment and get immediate rewards from feedback and The first comprehensive information collected ; Experience , , , Deposit into experience pool; Step 203, randomly obtain from the experience pool experience; enter Go to the TargetActor network to get the The allocation strategy for the execution of ; Step 204, calculate the first loss value and the second loss value , gradient descent method updates the Actor network and Critic network; ; ; ; in Represents input and The output of the Critic network is Indicates execution The comprehensive Q value when express and The output of the TargetCritic network; ; in, , , Represent the first, second and third balance factors respectively, represents the total proportion of the game at time step t, represents the average value of the total proportion of the game in the past n time steps, represents the expenditure proportion of the game at time step t, Indicates the preset target expenditure ratio. Indicates immediate reward; is the discount factor, which is a value between 0 and 1, used to balance the weights of immediate rewards and future rewards. The default value is 0.6; pass Update the Actor network by Update the Critic network; Step 205, every fixed number of network updates, update the TargetActor network so that its parameters are the same as those of the current Actor network, and update the TargetCritic network so that its parameters are the same as those of the current Critic network; Step 206, until the Actor network and the Critic network converge, the step is terminated, or after the number of loops of steps 203-205 reaches a set value, the Actor network replaces the allocation strategy model and the step is terminated.

[0015] In a preferred embodiment, an Internet marketing decision-making system based on deep learning optimization includes: Division module: It is used to divide each game into separate parts and set a separate total budget for each game; Data collection module: It is used to collect the comprehensive information of each game at n consecutive time points before the current time point; Data encoding module: used to generate No. 1 plane data for each game based on No. 1 comprehensive information encoding of n consecutive time points before and including the current time point of each game; Decision generation module: It is used to input the No. 1 plane data of each game into the allocation strategy model, and the allocation strategy model outputs the allocation decision of each game; Decision integration module: It is used to decode the allocation decisions of each game and integrate them to obtain a complete marketing decision.

[0016] The beneficial effects of the present invention are as follows: the present invention uses an Internet marketing decision-making method based on deep learning optimization to divide each game separately and set a separate model to output marketing decisions, and adds a method to determine whether the adjustment of the game will significantly affect the subsequent virtual currency marketing decisions. When it is determined that it will significantly affect the subsequent virtual currency marketing decisions, the game will be separated from the marketing decision system separately, and the execution of the marketing decisions of other games will not be affected. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of an Internet marketing decision-making method based on deep learning optimization of the present invention; Figure 2It is a flow chart of adjustment impact prediction in an Internet marketing decision-making method based on deep learning optimization of the present invention; Figure 3 It is a module schematic diagram of an Internet marketing decision-making system based on deep learning optimization of the present invention. DETAILED DESCRIPTION

[0018] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the present specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0019] At least one embodiment of the present invention discloses an Internet marketing decision-making method based on deep learning optimization, such as Figure 1 As shown, the following steps are included: Step 100, divide each game into separate categories and set a separate total budget for each game; Step 200, collecting the first comprehensive information of each game at n consecutive time points before the current time point, the first comprehensive information including: sales data, player data and market data; Sales data includes sales revenue, sales cost, and sales volume; Player data includes the number of players, activity, and payment rate; Market data includes market size, growth rate, and competition; In one embodiment of the present invention, the collected data is collected at fixed time intervals.

[0020] In one embodiment of the present invention, sales data is obtained through the enterprise's internal sales system, player data is obtained through communication with the game operation platform or developer, and market data is obtained through market research.

[0021] Step 300, generating No. 1 plane data for each game based on No. 1 comprehensive information encoding of a total of n consecutive time points before and including the current time point of each game; The first plane data includes a first bin for storing data, a first serial data stream indexed by a first bin to a first data object, and the first plane data also includes a first relationship list for storing association relationships between the first data objects indexed by the first bin, wherein the element value of the i-th row and j-th column of the first relationship list represents the quantitative value of the association relationship between the first data objects indexed by the i-th and j-th first bins; The association relationship includes high and low, corresponding to two quantitative values ​​respectively.

[0022] Data objects include players, sales, markets, and games; The data objects with high correlation are defined as: games and players in the games, games and the sales of their virtual currency, games and the market for the sales of their virtual currency; A No. 1 serial data stream of a No. 1 data object includes n No. 1 data units sorted in time, and the t-th No. 1 data unit represents the No. 1 comprehensive information collected at the t-th moment of the No. 1 data object.

[0023] In one embodiment of the present invention, the association between players and games is that the players belong to the game, the association between games and sales is that the players' play of the game will affect the sales of virtual currency, and the association between games and the market is that the players' play of the game will affect the sales of the entire market for virtual currency sales.

[0024] Step 400, inputting the No. 1 plane data of each game into an allocation strategy model, the allocation strategy model comprising a first encoder, a second encoder and a first fully connected layer, wherein the first encoder is used to input the No. 1 plane data, output a first identification feature to the second encoder, the second encoder outputs a second identification feature to the first fully connected layer, and the first fully connected layer outputs an allocation decision for each game; In one embodiment of the present invention, the first comprehensive information is input into the allocation strategy model after feature engineering; the first comprehensive information is in text mode, and the Word Embedding algorithm is used as a feature engineering method; In one embodiment of the present invention, the calculation formula of the first encoder is as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030] in represents the tth data unit of the first serial data stream corresponding to the vth bin of the first plane data, and They respectively represent the t-th and t-1-th first identification features of the No. 1 serial data stream corresponding to the v-th No. 1 bin of the No. 1 plane data, , , , and They represent the tth first, second, third, fourth and fifth intermediate features of the vth first face element of the first plane data, respectively. represents the t-1th fourth intermediate feature of the vth first face element of the first plane data, → represents the circulation direction, , , , , , , , Represents the first, second, third, fourth, fifth, sixth, seventh, and eighth serial recognition weight parameters respectively , , , Respectively represent the first, second, third, and fourth serial identification bias parameters represents the Hadamard product, , n represents the total number of data units No. 1 in serial data stream No. 1, when t=1 , tanh is the tanh function, Represents the Sigmoid function.

[0031] In one embodiment of the present invention, the calculation formula of the second encoder is as follows:

[0032] in represents the panel feature of the vth first panel in the lth layer, represents the set of No. 1 facets with high correlation with No. 1 facet v, represents the first plane recognition weight matrix of layer l, is the attention weight of the first facet u on the first facet v at the lth layer, which is calculated by the following formula:

[0033] in is a learnable attention vector, represents the vector concatenation operation, represents the second plane recognition weight matrix of layer l, represents transpose; and They represent the panel features of the vth first panel and the uth first panel in the l-1th layer respectively. , E represents the total number of layers, when l=1 , , when l=E is equal to the second identification feature of the first face element v, where and They respectively represent the t-th first identification feature of the No. 1 serial data stream corresponding to the v-th No. 1 bin and the u-th No. 1 bin of the No. 1 plane data, and n represents the total number of No. 1 data units of the No. 1 serial data stream.

[0034] In one embodiment of the present invention, the calculation formula of the first fully connected layer is as follows:

[0035] Where out represents a vector, and the components of the vector represent the total proportion and expenditure proportion respectively. represents the panel feature of the vth first panel in the Eth layer, represents the set of No. 1 facets with high correlation with No. 1 facet v, Represents the first full connection.

[0036] It should be noted that the total percentage indicates the percentage of the total budget of the game in the total budget of all games, and the expenditure percentage indicates the percentage of the actual expenditure on purchasing virtual currency for the game in the total budget of the game.

[0037] The allocation strategy model inputs the No. 1 plane data of one game at a time and outputs the marketing decision for that game.

[0038] Step 500, decoding and integrating the allocation decisions of each game to obtain a complete marketing decision.

[0039] Define the game that needs to be adjusted when executing the marketing decision as Game A. Use the following steps to determine the impact of the adjustment of Game A on the marketing of virtual currency, such as Figure 2 As shown, including: Step 101, collecting the second comprehensive information of the game A at n consecutive time points before the current time point, the second comprehensive information including: adjustment information, player data, player feedback information, and social media information; Adjustment information includes: update logs, virtual currency adjustment information; Step 102, generating a second serial data stream based on the second comprehensive information encoding of n consecutive time points before and including the current time point, wherein the second serial data stream of a second data object comprises n second data units sorted in time, and the t-th second data unit represents the second comprehensive information collected at the t-th moment of the second data object; Step 103, inputting the second serial data stream into the impact prediction model, the impact prediction model comprising a third encoder and a second fully connected layer, wherein the third encoder is used to input the second serial data stream, output the third identification feature to the second fully connected layer, and the second fully connected layer outputs the impact level; In one embodiment of the present invention, the second comprehensive information is input into the allocation strategy model after feature engineering; the first comprehensive information is in text mode, and the Word Embedding algorithm is used as a feature engineering method; In one embodiment of the present invention, the calculation formula of the third encoder is as follows:

[0040]

[0041]

[0042] in represents the tth third identification feature, represents the t-1th third identification feature, and denote the first and second intermediate features respectively, represents the tth data unit No. 2 of the No. 2 serial data stream, and is the serial recognition weight matrix, is the serial identification bias parameter, Represents the Sigmoid function.

[0043] In one embodiment of the present invention, the calculation formula of the second fully connected layer is as follows:

[0044] in Represents a vector, the components of which represent the probability of each impact level of game A at the next time point, and the impact level with the highest probability is taken as the output. represents the tth third identification feature, Represents the second full connection.

[0045] In one embodiment of the present invention, the impact levels are divided into three levels: no impact, slight impact, and severe impact.

[0046] Step 104: determine whether it is necessary to separate the game from the marketing decision system based on the outputted predicted impact level of the game.

[0047] In one embodiment of the present invention, when the predicted impact level of a game adjustment is determined to be a severe impact, it is considered that the adjustment has seriously affected the transaction of the game's virtual currency, and the game is separated from the marketing decision system.

[0048] An Internet marketing decision-making system based on deep learning optimization, such as Figure 3 As shown, including: Division module M1: used to divide each game into separate parts and set a separate total budget for each game; Data collection module M2: used to collect comprehensive information of n consecutive time points before the current time point of each game; Data encoding module M3: used to generate No. 1 plane data for each game based on No. 1 comprehensive information encoding of n consecutive time points before and including the current time point of each game; Decision generation module M4: used to input the No. 1 plane data of each game into the allocation strategy model, and the allocation strategy model outputs the allocation decision of each game; Decision integration module M5: It is used to decode and integrate the allocation decisions of each game to obtain a complete marketing decision.

[0049] In one embodiment of the present invention, the allocation strategy model is trained using a reinforcement learning method, and the training steps of the allocation strategy model are as follows: Step 201, copy the allocation strategy model to generate the Actor network and the TargetActor network, use the feedforward neural network as the Critic network, copy the Critic network to obtain the TargetCritic network, and randomly initialize the parameters of the Actor network, the TargetActor network, the Critic network, and the TargetCritic network; Step 202, input The first comprehensive information collected Go to the Actor network to get the The allocation strategy for the execution of ; No. The allocation strategy for the execution of Add random noise to get The noise allocation strategy ; The first The allocation strategy for the execution of Input into the environment and get immediate rewards from feedback and The first comprehensive information collected ; Experience , , , Deposit into experience pool; Step 203, randomly obtain from the experience pool experience; enter Go to the TargetActor network to get the The allocation strategy for the execution of ; Step 204, calculate the first loss value and the second loss value , gradient descent method updates the Actor network and Critic network; ; ; ; in Represents input and The output of the Critic network is Indicates execution The comprehensive Q value when express and The output of the TargetCritic network; ; in, , , Represent the first, second and third balance factors respectively, represents the total proportion of the game at time step t, represents the average value of the total proportion of the game in the past n time steps, represents the expenditure proportion of the game at time step t, Indicates the preset target expenditure ratio. Indicates immediate reward; is the discount factor, which is a value between 0 and 1, used to balance the weights of immediate rewards and future rewards. The default value is 0.6; pass Update the Actor network by Update the Critic network; Step 205, every fixed number of network updates, update the TargetActor network so that its parameters are the same as those of the current Actor network, and update the TargetCritic network so that its parameters are the same as those of the current Critic network; Step 206, until the Actor network and the Critic network converge, the step is terminated, or after the number of loops of steps 203-205 reaches a set value, the Actor network replaces the allocation strategy model and the step is terminated.

[0050] In one embodiment of the present invention, the impact prediction model uses a multi-classification cross entropy loss function as the loss function of the model, and the calculation formula is as follows:

[0051] in represents the loss value, M represents the total number of classification categories, represents the total number of training samples, It is The symbol value of the training sample, if the The true classification of the training sample is the cth classification category and its value is 1, otherwise it is 0. It is the first factor that affects the prediction model. The probability that a training sample belongs to the cth classification category, Represents the logarithmic function with the natural constant e as the base.

[0052] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. An Internet marketing decision-making method based on deep learning optimization, characterized in that: The following steps are involved: Step 100, divide each game into separate categories and set a separate total budget for each game; Step 200, collecting the first comprehensive information of each game at n consecutive time points before the current time point, the first comprehensive information including: sales data, player data and market data; The data is collected at fixed time intervals; Step 300, generating No. 1 plane data for each game based on No. 1 comprehensive information encoding of a total of n consecutive time points before and including the current time point of each game; Step 400, inputting the No. 1 plane data of each game into an allocation strategy model, the allocation strategy model comprising a first encoder, a second encoder and a first fully connected layer, wherein the first encoder is used to input the No. 1 plane data, output a first identification feature to the second encoder, the second encoder outputs a second identification feature to the first fully connected layer, and the first fully connected layer outputs an allocation decision for each game; Step 500, decoding and integrating the allocation decisions of each game to obtain a complete marketing decision.

2. The Internet marketing decision-making method based on deep learning optimization according to claim 1 is characterized in that: Sales data includes sales revenue, sales cost, and sales volume; Player data includes the number of players, activity, and payment rate; Market data includes market size, growth rate, and competition.

3. The Internet marketing decision-making method based on deep learning optimization according to claim 1 is characterized in that: The first plane data includes a first bin for storing data, a first bin indexing a first serial data stream to a first data object, and the first plane data also includes a first relationship list for storing association relationships between the first data objects indexed by the first bin, wherein the element value of the i-th row and j-th column of the first relationship list represents the quantitative value of the association relationship between the first data objects indexed by the i-th and j-th first bins; The association relationships include high and low, corresponding to two quantitative values; Data objects include players, sales, markets, and games; The data objects with high correlation are defined as: games and players in the games, games and the sales of their virtual currency, games and the market for the sales of their virtual currency; A No. 1 serial data stream of a No. 1 data object includes n No. 1 data units sorted in time, and the t-th No. 1 data unit represents the No. 1 comprehensive information collected at the t-th moment of the No. 1 data object.

4. The Internet marketing decision-making method based on deep learning optimization according to claim 1 is characterized in that: The output of the allocation strategy model represents a vector, and the components of the vector represent the total proportion and expenditure proportion respectively. The total proportion represents the proportion of the total budget of the game to the total budget of all games, and the expenditure proportion represents the proportion of the actual investment in purchasing virtual currency for the game to the total budget of the game.

5. The Internet marketing decision-making method based on deep learning optimization according to claim 1 is characterized in that: The allocation strategy model is trained using the reinforcement learning method. The training steps of the allocation strategy model are as follows: Step 201, copy the allocation strategy model to generate the Actor network and the TargetActor network, use the feedforward neural network as the Critic network, copy the Critic network to obtain the TargetCritic network, and randomly initialize the parameters of the Actor network, the TargetActor network, the Critic network, and the TargetCritic network; Step 202, input The first comprehensive information collected Go to the Actor network to get the The allocation strategy for the execution of ; No. The allocation strategy for the execution of Add random noise to get The noise allocation strategy ; The first The allocation strategy for the execution of Input into the environment and get immediate rewards from feedback and The first comprehensive information collected ; Experience , , , Deposit into experience pool; Step 203, randomly obtain from the experience pool experience; enter Go to the TargetActor network to get the The allocation strategy for the execution of ; Step 204, calculate the first loss value and the second loss value , gradient descent method updates the Actor network and Critic network; ； ； ； in Represents input and The output of the Critic network is Indicates execution The comprehensive Q value when express and The output of the TargetCritic network; ； in, , , Represent the first, second and third balance factors respectively, represents the total proportion of the game at time step t, represents the average value of the total proportion of the game in the past n time steps, represents the expenditure proportion of the game at time step t, Indicates the preset target expenditure ratio. Indicates immediate reward; is the discount factor, which is a value between 0 and 1, used to balance the weights of immediate rewards and future rewards. The default value is 0.6; pass Update the Actor network by Update the Critic network; Step 205, every fixed number of network updates, update the TargetActor network so that its parameters are the same as those of the current Actor network, and update the TargetCritic network so that its parameters are the same as those of the current Critic network; Step 206, until the Actor network and the Critic network converge, the step is terminated, or after the number of loops of steps 203-205 reaches a set value, the Actor network replaces the allocation strategy model and the step is terminated.

6. An Internet marketing decision-making method based on deep learning optimization, characterized in that: Define the game that needs to be adjusted when executing the marketing decision as Game A. The following steps are used to determine the impact of the adjustment of Game A on the marketing of virtual currency, including: Step 101, collecting the second comprehensive information of the game A at n consecutive time points before the current time point, the second comprehensive information including: adjustment information, player data, player feedback information, and social media information; Adjustment information includes: update logs, virtual currency adjustment information; Step 102, generating a second serial data stream based on the second comprehensive information encoding of n consecutive time points before and including the current time point, wherein the second serial data stream of a second data object comprises n second data units sorted in time, and the t-th second data unit represents the second comprehensive information collected at the t-th moment of the second data object; Step 103, inputting the second serial data stream into the impact prediction model, the impact prediction model comprising a third encoder and a second fully connected layer, wherein the third encoder is used to input the second serial data stream, output the third identification feature to the second fully connected layer, and the second fully connected layer outputs the impact level; Step 104: determine whether it is necessary to separate the game from the marketing decision system based on the outputted predicted impact level of the game.

7. The Internet marketing decision-making method based on deep learning optimization according to claim 6 is characterized in that: The output of the impact prediction model is represented as a vector, the components of which represent the probability of each impact level of game A at the next time point, and the impact level with the highest probability is taken as the output.

8. The Internet marketing decision-making method based on deep learning optimization according to claim 7 is characterized in that: After feature engineering, the second comprehensive information is input into the impact prediction model; the second comprehensive information includes text modality, and the Word Embedding algorithm is used as a feature engineering method.

9. The Internet marketing decision-making method based on deep learning optimization according to claim 8 is characterized in that: The impact prediction model uses the multi-classification cross entropy loss function as the loss function of the model. The calculation formula is as follows: ； in represents the loss value, M represents the total number of classification categories, represents the total number of training samples, It is The symbol value of the training sample, if the The true classification of the training sample is the cth classification category and its value is 1, otherwise it is 0. It is the first factor that affects the prediction model. The probability that a training sample belongs to the cth classification category, Represents the logarithmic function with the natural constant e as the base.

10. An Internet marketing decision-making system based on deep learning optimization, characterized in that: include: Division module: It is used to divide each game into separate parts and set a separate total budget for each game; Data collection module: It is used to collect the comprehensive information of each game at n consecutive time points before the current time point; Data encoding module: used to generate No. 1 plane data for each game based on No. 1 comprehensive information encoding of n consecutive time points before and including the current time point of each game; Decision generation module: It is used to input the No. 1 plane data of each game into the allocation strategy model, and the allocation strategy model outputs the allocation decision of each game; Decision integration module: It is used to decode the allocation decisions of each game and integrate them to obtain a complete marketing decision.

Images