A method, system, electronic device and storage medium for generating an advertising placement plan

Abstract

Embodiments of the present application disclose a method, system, electronic device, and storage medium for generating an advertising placement plan, including: training and generating an adversarial network time series model based on first historical advertising placement combination sequence data, preset parameters, and a random vector; inputting second historical advertising placement combination sequence data and target parameters into the generated adversarial network time series model to obtain a simulated target advertising placement combination; processing the simulated target advertising placement combination to determine a target advertising placement plan. It can automatically generate a more scientific advertising placement plan for cross-platform advertising placement, replacing the previous technical solution of advertising placement based on human experience, which can reduce labor consumption, improve decision-making efficiency, facilitate a more scientific and reasonable budget allocation, and optimize the specific advertising placement situation for each future day.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of advertising placement, for example, to a method, system, electronic device and storage medium for generating an advertising placement plan. Background Art

[0002] Due to information islands and data barriers between platforms, advertisers have long faced the problems of how to make scientific budget allocation and cross-platform resource placement.

[0003] Related technologies can only solve the placement plan problems within a single platform. Based on statistical or machine learning models, they summarize and analyze user behavior patterns to plan new placement plans, and can achieve precise placement within platforms with data interconnection. However, there are data barriers between different platforms. For cross-platform placement, existing solutions stay at the stage of effect evaluation and experience-based placement. They obtain a general optimization direction based on the effect evaluation results and then cooperate with industry experience for placement. Due to relying on human experience, the decision-making cycle is long, the placement efficiency is low, and at the same time, there is a lack of specific daily placement scientific basis and effect controllability, resulting in budget waste. Summary of the Invention

[0004] The embodiments of the present application provide a method, system, electronic device and storage medium for generating an advertising placement plan to improve the situation in related technologies where it is impossible to reasonably formulate an advertising placement plan for cross-platform scenarios.

[0005] In a first aspect, the embodiments of the present application provide a method for generating an advertising placement plan, including:

[0006] Training and generating an adversarial network time series model according to the first historical advertising placement combination sequence data, preset parameters and a random vector;

[0007] Inputting the second historical advertising placement combination sequence data and target parameters into the generated adversarial network time series model to obtain a simulated target advertising placement combination;

[0008] Processing the simulated target advertising placement combination to determine a target advertising placement plan.

[0009] In a second aspect, the embodiments of the present application provide a system for generating an advertising placement plan, including:

[0010] A front-end module, a business processing module, and a data processing module;

[0011] The front-end module is configured to send a request for generating a target advertising placement plan and display the target advertising placement plan;

[0012] The business processing module is configured to receive the request for generating the target advertising placement plan, input training data to the data processing layer, and generate an adversarial network time series model;

[0013] The data processing module is configured to receive the training data and the adversarial network time series model, obtain the target advertising placement plan, and return the target advertising placement plan to the business processing layer.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, which includes: one or more processors;

[0015] a memory configured to store one or more programs,

[0016] When the one or more programs are executed by the one or more processors, the one or more processors implement the advertising placement plan generation method according to any embodiment of the present application.

[0017] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the advertising placement plan generation method according to any embodiment of the present application is implemented.

[0018] In an embodiment of the present invention, an adversarial network time series model is trained based on the first historical advertising placement combination sequence data, preset parameters, and a random vector; the second historical advertising placement combination sequence data and target parameters are input into the adversarial network time series model to obtain a simulated target advertising placement combination; the simulated target advertising placement combination is processed to determine the target advertising placement plan. It can automatically generate a more scientific advertising placement plan for cross-platform advertising placement, replacing the previous technical solution of advertising placement based on human experience, which can reduce labor consumption, improve decision-making efficiency, facilitate a more scientific and reasonable budget allocation, and optimize the specific advertising placement situation in the future every day. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of steps of an advertising placement plan generation method provided by an embodiment of the present application;

[0020] Figure 2 is an architecture diagram of joint training of an adversarial network time series model provided by an embodiment of the present application;

[0021] Figure 3 is a flowchart of data processing by an adversarial network time series model provided by an embodiment of the present application;

[0022] Figure 4It is a flowchart showing a prediction model in a generative adversarial network time series model verification module provided by an embodiment of the present application for processing data;

[0023] Figure 5 It is a structural block diagram of an advertising placement plan generation system provided by an embodiment of the present application;

[0024] Figure 6 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0025] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the example embodiments described herein are only for explaining the present application and not for limiting the present application. Additionally, it should be noted that for the sake of description, only parts related to the present application are shown in the drawings rather than all the structures. Without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0026] Figure 1 It is a step flowchart of an advertising placement plan generation method provided by an embodiment of the present application. The embodiments of the present application are applicable to the situation of generating an advertising placement plan for cross-platform scenarios. This method can be executed by the advertising placement plan generation system of the embodiments of the present application. The advertising placement plan generation system can be implemented by hardware or software and integrated in the electronic device provided by the embodiments of the present application. In one embodiment, as Figure 1 shown, the advertising placement plan generation method of the embodiments of the present application may include the following steps:

[0027] S110. Train a generative adversarial network time series model according to the first historical advertising placement combination sequence data, preset parameters, and a random vector.

[0028] In the embodiments of the present application, the first historical advertising placement combination sequence data refers to the historical data of a certain advertising resource being placed on different platforms within a certain time period. The preset parameters refer to the return data of the historical advertising placement combination sequence data.

[0029] In one embodiment, training the generative adversarial network time series model according to the first historical advertising placement combination sequence data and preset parameters includes: training the representation network in the generative adversarial network time series model according to the first historical advertising placement combination sequence data and the preset parameters; training the adversarial network in the generative adversarial network time series model according to the random vector; training the supervision network according to the representation network and the adversarial network; wherein, the representation network, the adversarial network, and the supervision network constitute the generative adversarial network time series model.

[0030] In one embodiment, training the representation network in the generative adversarial network time series model according to the first historical advertisement placement combination sequence data and the preset parameters includes: determining a first placement feature (S, X 1:T ) of the first historical advertisement placement combination sequence data, and mapping the first placement feature (S, X 1:T ) from the original space to the latent space through an embedding network to obtain a corresponding first latent vector h S , h 1:T = e(s, x 1:t ); restoring the first latent vector h S , h 1:T = e(s, x 1:t ) to a second placement feature in the original space wherein, the representation network includes the embedding network and the restoration network. For example, the embedding network maps the placement feature to the latent space, enabling the adversarial network to learn the potential temporal dynamics of the placement combination data through a low-dimensional representation. Let H S , H X represent the latent vector spaces corresponding to the feature spaces S, X, and the embedding function e maps the static and temporal features to their latent spaces h S , h 1 : T = e(s, X 1 : t): h S = e S (s)h t = e x (h S , h t-1 , x t ), where e S and e X are the embedding networks for static features and temporal features respectively. The restoration network provides a mapping from the latent space to the original space, and r restores the static and temporal latent variables to their original feature expressions: We implement it through a feed-forward network: The embedding and restoration functions have various optional structures, which are autoregressive and the output of each step can only depend on the previous information.

[0031] In one embodiment, the adversarial network includes the sequence generation network and the sequence discrimination network. Training the adversarial network in the generative adversarial network time series model according to the random vector includes: the sequence generation network generating a second latent vector feature h S , h 1：T = g(z S , z 1：T ) in the latent space according to the random vector; the sequence discrimination network discriminating the second latent vector feature h S , h1：T = g(z S , z 1：T ) to make a discrimination and obtain the sequence classification y S , y 1：T = d(h S , h 1：T ); where h S = g S (z S ), h t = g X (h S , h t-1 , z t ).

[0032] For example, Generative Adversarial Networks (GANs) include a sequence generation network and a sequence discrimination network. The sequence generation network uses a large number of random vectors generated by a known simple distribution in the latent space as inputs and generates a second latent vector feature h S , H X in the latent space H S , h 1：T = g(z S , Z 1：T ): h S = g S (z S ), h t = g X (h S , h t-1 , z t ). Where g S : Z S → H S is a static feature generator, and g X : H S × H S × Z X → H X is a temporal feature recurrent generator. The random vector z S can be sampled from the distribution, and z t then follows a random process. The sequence discrimination network is set to discriminate the probability of the authenticity of the second latent vector feature. The discrimination network d: H S × Π t X → [0, 1] × Π t [0, 1] accepts static and temporal features and returns the sequence classification y S , y 1：T = d(h S , h 1：T ). The discriminator is implemented using a bidirectional recurrent network plus a feedforward output layer: y S = dS (h S ) wherein respectively represent the forward and backward hidden state sequences, is the recurrent equation.d S , d X is the output of the classification function. Similarly, the generator is autoregressive and there are various optional architectures.

[0033] In one embodiment, the first hidden vector h S , h 1：T = e(s, x 1：t ) is restored to the second projection feature in the original space After that, it further includes: calculating the reconstruction loss

[0034] In one embodiment, the sequence generation network generates the next h S , h 1 : t - 1 by receiving the artificial embedding h t , and calculates the gradient through the unsupervised loss; according to the sequence classification y S , y 1：T = d(h S , h 1：T ) calculates the unsupervised loss According to the real data h S , h 1：T and the calculated gradient, calculates the supervised loss L S = E s，x1：T～p [∑ t ||h t - g X (h S , h t-1 , z t )|| 2 .

[0035] S120. Input the second historical advertisement placement combination sequence data and the target parameters into the generative adversarial network time series model to obtain the simulated target advertisement placement combination.

[0036] The second historical advertising placement portfolio sequence data is different from the first historical advertising placement portfolio sequence data used to train the generative adversarial network time series model in S110. The second historical advertising placement portfolio sequence data is the historical placement sequence data within a certain time period selected by any user according to their actual situation when obtaining the target advertising placement plan using the advertising placement plan generation system. The target parameters include the average unit price of advertising on each platform when the user generates a new advertising placement plan, the budget threshold for advertising placement, and the combined duration of advertising placement on each platform, etc. The simulated target advertising placement portfolio is the new advertising placement portfolio and return data obtained by inputting the second historical advertising placement portfolio sequence data and the target parameters input by the user into the trained generative adversarial network time series model.

[0037] In one embodiment, the step of inputting the second historical advertising placement portfolio sequence data and the target parameters into the generative adversarial network time series model to obtain the simulated target advertising placement portfolio includes: inputting the second historical advertising placement portfolio sequence data and the synthetic data into the representation network, and inputting the synthetic data of the second historical advertising placement portfolio sequence data into the adversarial network for joint training and iterating more than a preset threshold to generate the simulated target advertising placement portfolio data.

[0038] S130. Process the simulated target advertising placement portfolio to determine the target advertising placement plan.

[0039] In the embodiment of the present application, the simulated target advertising placement portfolio is not the final target advertising placement plan. After the generative adversarial network time series model outputs the simulated target advertising placement portfolio, it is necessary to perform inspection processing and search processing on the simulated target advertising placement portfolio. The inspection processing uses two types of sequence prediction models, ConvLSTM (as shown in Figure 4 ) and Transformer (the specific data processing process of such models is not shown in the drawings of the present application), to test the authenticity of the simulated target advertising placement portfolio data. On the basis of confirming the authenticity of the simulated target advertising placement portfolio data, then perform search processing on the simulated target advertising placement portfolio. The search processing refers to traversing the combined return data of all simulated target advertising placement portfolios, and according to the target parameters input by the user, such as the average unit price of advertising resources on each platform, calculating whether the budget of the advertising placement plan is within the budget threshold range set by the user, and according to the combined duration of advertising resource placement on each platform input by the user, searching for the optimal advertising placement portfolio and its corresponding budget allocation plan.

[0040] In one embodiment, the processing of the simulated target advertising placement portfolio to determine the target advertising placement plan includes: inspecting the simulated target advertising placement portfolio to determine whether it is authentic; and searching the simulated target advertising placement portfolio to output the optimal solution therein.

[0041] The method for generating an advertising placement plan according to an embodiment of the present application includes training a generative adversarial network time series model based on the first historical advertising placement portfolio sequence data, preset parameters, and a random vector; inputting the second historical advertising placement portfolio sequence data and target parameters into the generative adversarial network time series model to obtain a simulated target advertising placement portfolio; and processing the simulated target advertising placement portfolio to determine the target advertising placement plan. This can automatically generate a more scientific advertising placement plan for cross-platform advertising placement, replacing the previous technical solution of advertising placement based on human experience, reducing labor consumption, improving decision-making efficiency, facilitating a more scientific and reasonable budget allocation, and optimizing the specific advertising placement situation for each future day.

[0042] Figure 2 FIG. is an architecture diagram of the joint training of a generative adversarial network time series model provided by an embodiment of the present application. As Figure 2 shown, the main process of the joint training of the generative adversarial network time series model according to the embodiment of the present application is as follows:

[0043] For the resource placement sequence data, the variable dimension includes a time variable, the placement data variables for each platform, and the final return variable. Each piece of data includes two types of features: static features (not changing with time, such as the minimum number of visits to Weibo) and time features (changing with time, such as the number of visits changing with the increase of the topic). Let S be the vector space of static features, X be the vector space with time series features, and let S∈S, X∈X be random vectors in the corresponding spaces. Assume that the placement feature is a tuple in the form of (S, X 1：T ) that satisfies a certain joint distribution p. The training set is denoted as The goal is to make the training data D learn the distribution and be closest to the true distribution p(S, X 1：T ). Additionally, by using the autoregressive decomposition of the distribution p(S, X 1：T ) = p(S)Π t p(X t |S, X 1：T-1 ), the distribution is learned to be closest to the distribution p(X t |S, X 1：T-1 ). The final goal can be described as two minimization problems,

[0044] 1) Minimize the distribution difference:

[0045]

[0046] D is to calculate the gap between two distributions.

[0047] 2) Minimize the difference in conditional probability distributions:

[0048]

[0049] The generative adversarial network time series model adopted in the embodiments of this application consists of four networks in total: an embedding network, a recovery network, a sequence generation network, and a sequence discriminator network. The embedding network maps the placement combination feature sequences S, X to the latent space H S , H X , capturing the static and dynamic features of the placement feature sequences. The recovery network then restores the data from the latent space to the original space The sequence generation network and the sequence discriminator network together constitute an adversarial network and perform operations in the latent space:

[0050] 1) The potential dynamics of real data and synthetic data are constrained by the reconstruction loss and the supervised loss:

[0051]

[0052] where λ ≥ 0 is a hyperparameter that balances the two losses.

[0053] 2) The training of the adversarial network includes a generator and a discriminator, and is completed by minimizing the unsupervised loss and maximizing the supervised loss:

[0054]

[0055] where η ≥ 0 is another hyperparameter that balances the two losses. L S The included embedding process not only helps reduce the dimension of the learning space, but also helps the generator learn temporal relationships from the data.

[0056] The four networks are jointly trained and iterated more than 10,000 times, so that the finally generated placement combination data is "similar" enough to the real placement combination, and the distribution is sufficient to cover the real data distribution.

[0057] The output module outputs a large number of simulated placement combinations to the test module and the search module, and respectively conducts prediction effect tests and searches for the optimal solution for the generated placement combinations.

[0058] The inspection module of this application includes two types of sequence prediction models, ConvLSTM and Transformer, to test the prediction scores of the generated data: Mean Square Error (MAPE), Root Mean Square Error (RMSE), and goodness of fit R-squared.

[0059] Figure 3 It is a flowchart of processing data by a generative adversarial network time series model provided by an embodiment of this application. As Figure 3 shown, the main process of processing data by the generative adversarial network time series model in the embodiment of this application is as follows:

[0060] The model jointly embeds the network and the adversarial network to learn together. The learning is mainly responsible for optimizing two loss functions:

[0061] 1) The reconstruction-generated loss loss:

[0062]

[0063] 2) In the adversarial network, the generator generates the next h S , h 1：t-1 by receiving the artificial embedding h t· Then, the gradient is calculated through the unsupervised loss. Then, by providing the true data h S , h 1：T and the accurate classification y of the artificial data S , y 1：T to maximize (discriminator) or minimize (generator) the likelihood to improve the classification ability.

[0064]

[0065] Outside the adversarial feedback between the generator and the discriminator, additional losses are introduced to constrain the learning and encourage the generator to capture the sequential conditional distribution in the data. During training, the generator receives the embedding sequence of the true data h 1：t-1 to generate the next latent vector h t . The gradient can now be calculated on the loss of the difference between the distributions p(H t |H S , H 1：t-1 ) and . The supervised loss generated using maximum likelihood:

[0066] L S = E s，x1：T～p [∑ t ||h t - g X (hS , h t-1 , z t )|| 2

[0067] where g X (h S , h t-1 , z t ) uses a sample z t to approximate L U When promoting the generator to create a sequence, LS further ensures that it produces a similar step-by-step conversion effect.

[0068] Advertising placement plan generation method process:

[0069] First step, the user uploads sequence data, and economic indicators and macro variables can be added according to the actual scenario. Set the time length of the placement combination to be generated.

[0070] Second step, model operation: The model first performs embedding encoding on the data, and then performs joint adversarial learning to generate time series data.

[0071] Figure 3 The execution process after the model receives the data: Receive the model training data and parameters, first select the normalization method according to the parameters; then construct each supervised training data according to the step size parameter; train the supervised generation network with the data obtained by mapping the real input; then jointly train the entire network with adversarial, supervised, and recovery losses; finally, by sampling a large number of samples from the known distribution, constrain it to the real investment sequence data through the upload network. Many corresponding operations can be performed on the output of the model to meet business requirements, including verifying the authenticity of the generated sequence, searching for the sequence with the highest sales, visually displaying the final result, and saving the result, etc.

[0072] Figure 4 It is the flowchart of the data processing of the prediction model included in a generative adversarial network time series model verification module provided by an embodiment of the present application. The figure shows the data processing process of the output return prediction of the ConvLSTM model.

[0073] Figure 5 It is the structural block diagram of an advertising placement plan generation system provided by an embodiment of the present application. As Figure 5 shown, the advertising placement plan generation system of the embodiment of the present application includes: a front-end module, a business processing module, and a data processing module;

[0074] The front-end module is set to send a request to generate a target advertising placement plan and display the target advertising placement plan;

[0075] ​The business processing module is configured to receive the request for generating the target advertising placement plan, input training data into the data processing layer, and generate an adversarial network time series model;

[0076] The data processing module is configured to receive the training data and the adversarial network time series model, obtain the target advertising placement plan, and return the target advertising placement plan to the business processing layer.

[0077] The system consists of three layers, namely the front-end layer, the business logic layer, and the data processing layer. The three layers jointly implement operations such as business creation, execution, saving, and verification. The specific description is as follows:

[0078] The main functions of the front-end layer are: user registration and login, uploading data, displaying data / business, managing business, viewing / downloading results, and initializing data processing. First, the user needs to register an account and then log in; if creating a new task, the user needs to click "Create Task", then upload file, model parameters and other information according to the prompts, click "Submit", and then upload the data and parameters to the background through the business API of the created task, and then return to the home page with the business status being "executing". When the business execution is completed, the server returns the execution status, and the front-end updates the execution status. Clicking "Display" can display the model results, including the generated sequences with the highest predictions for each channel and the corresponding investment suggestions. The model results can be saved by clicking "Download".

[0079] The business logic layer is responsible for receiving business requests, verifying identities, saving tasks and data, calling models, obtaining model results, and analyzing the returned results to return task results. When the business logic layer receives a task request from the front-end, according to the request type, it first verifies the user's identity and then performs corresponding processing; if it is to create and execute a task, after successful identity verification, it first saves the model data and parameters to the database, and at the same time associates the task creation, management, and user accounts; then uses the model and parameters through the model call API, and saves the task status; when receiving the model results, it modifies the task status and at the same time notifies the front-end that the task status has changed. If it is other operations on the task, such as deleting or modifying the users, etc., it can directly perform the operations and return the execution results.

[0080] The data processing layer is mainly responsible for receiving the training data input by the business logic layer, training the model, saving the model results, and returning the model results to the business logic layer. The model receives the data and parameters transmitted by the business logic layer, uses the data and parameters for model training; saves the model output results to the database; returns the result status of the model processing to the business logic layer.

[0081] Referring to Figure 6 , a schematic structural diagram of an electronic device in an example of the present application is shown. AsFigure 6 As shown in Figure 6 , the electronic device may include: a processor 801, a memory 802, a display screen 803 with touch function, an input device 804, an output device 805, and a communication device 806. The number of processors 801 in the device may be one or more. Figure 6 Here, one processor 801 is taken as an example. The processor 801, memory 802, display screen 803, input device 804, output device 805, and communication device 806 of the device may be connected via a bus or other means. Figure 6 Here, connection via a bus is taken as an example. The device is configured to execute the advertising placement plan generation method provided in any embodiment of the present application.

[0082] An embodiment of the present application also provides a computer-readable storage medium. When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can execute the advertising placement plan generation method described in the above method embodiment.

[0083] It should be noted that for the embodiments of the electronic device and the storage medium, since they are basically similar to the method embodiments, the description is relatively simple. For related parts, refer to the partial description of the method embodiments.

[0084] The present application applies the generative adversarial network in image generation technology to the field of placement plan / budget allocation. Compared with the traditional placement plan / budget allocation method, the present application adopts a generative adversarial network time series model, uses the long-term resource placement combination data as a feature matrix, and the encoder uses a large number of long-term resource placement feature tensors as training input data. The Time Series Generative Adversarial Network (TimeGan) adds supervised learning and an embedding network on the basis of the general generative adversarial network. Therefore, the present application utilizes the flexibility of unsupervised learning of the generative adversarial network and the ability of the supervised network to capture temporal dynamics to generate diverse sequence data that can simulate human placement styles and characteristics, and can more scientifically and efficiently guide future placements. By jointly training supervised learning and the adversarial network to capture the progressive conditional distribution of real data, the generated new sequences can inherit the temporal dynamics of the real data.

[0085] The present application provides an advertising placement plan generation system, which is divided into three layers: the underlying data operation layer, the business logic layer, and the front-end application layer. The user inputs historical placement and effect return data at the front end. The data is processed by the business layer and transmitted to the operation layer. After the generation model is trained, the results are returned to the front end. The front end displays the optimal placement combination and estimated effect return for the future time (e.g., the next month). The background calculates according to the latest average unit price of resources and displays the corresponding budget allocation plan. After creating a project, the user can select the type of model to be used and view the final result with one click.

[0086] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0087] Note that the above are only example embodiments of this application and the technical principles applied. Those skilled in the art will understand that this application is not limited to the specific embodiments described here, and various obvious transformations, re-adjustments, and substitutions can be made by those skilled in the art without departing from the protection scope of this application. Therefore, although this application has been described in more detail through the above embodiments, this application is not limited to the above embodiments only. Without departing from the concept of this application, more other equivalent embodiments can be included, and the scope of this application is determined by the scope of the appended claims.

Claims

1. An advertising placement plan generation method, including: Train a generative adversarial network time series model based on the first historical advertising placement portfolio sequence data, preset parameters, and random vectors, including: training the representation network in the generative adversarial network time series model according to the first historical advertising placement portfolio sequence data and the preset parameters: determining the first placement feature (S, X 1:T ) of the first historical advertising placement portfolio sequence data, and mapping the first placement feature (S, X 1:T ) from the original space to the latent space through an embedding network to obtain the corresponding first latent vector h S , h 1:T = e(s, x 1:t ), S represents the vector space of static features, X 1:T represents the vector space with time series features s ∈ S, x 1:T ∈ X 1:T , e() is the embedding function representing the embedding network, and restoring the first latent vector h S , h 1:T = e(s, x 1:t ) to the second placement feature in the original space r() is the restoration function representing the restoration network, and the representation network includes the embedding network and the restoration network; training the adversarial network in the generative adversarial network time series model according to the random vector: the sequence generation network generates a second latent vector feature h S , h 1:T = g(z S , z 1:T ) in the latent space according to the random vector, z S represents a random vector sampled from a distribution, z t follows a random process, g() is the function representing the sequence generation network, and the sequence discriminator network discriminates the second latent vector feature h S , h 1:T = g(z S , z 1:T ) to obtain the sequence classification y S , y 1:T = d(h S , h 1:T ), h S = g S (Z S ), h t = g X (h S , h t-1 , Z t ), d() is the function representing the sequence discriminator network, g S : Z S → H S is a static feature generator, g X :H S ×H S ×Z X →H X is a sequential feature cycle generator, H S 、H X represent the corresponding latent vector spaces of the feature spaces S and X 1:T The adversarial network includes the sequence generation network and the sequence discrimination network. The supervision network is trained according to the representation network and the adversarial network. Among them, the representation network, the adversarial network, and the supervision network constitute the generative adversarial network time series model; Inputting the second historical advertising placement portfolio sequence data and target parameters into the generative adversarial network time series model to obtain a simulated target advertising placement portfolio, including: inputting the second historical advertising placement portfolio sequence data and synthetic data into the representation network, inputting the synthetic data of the second historical advertising placement portfolio sequence data into the adversarial network for joint training and iterating more than a preset threshold to generate simulated target advertising placement portfolio data; Processing the simulated target advertising placement portfolio to determine the target advertising placement plan, including: inspecting the simulated target advertising placement portfolio to determine whether the simulated target advertising placement portfolio is real, and searching the simulated target advertising placement portfolio to output the optimal solution therein.

2. The method according to claim 1, wherein, Restore the first hidden vector h S , h 1:T = e(s, x 1:t ) to the second projection feature in the original space After that, it further includes: Calculate the reconstruction loss 3. The method according to claim 1, wherein, The sequence generation network generates the next h by receiving artificial embeddings h S , h 1:t-1 and calculates the gradient through an unsupervised loss; t ​ Classify y according to the said sequence S , y 1:T = d(h S , h 1:T ) Calculate the unsupervised loss Based on the real data h S , h 1:T and the calculated gradient, calculate the supervised loss L S = E S,x1:T～p [∑ t ||h t - g X (h S , h t-1 , z t )|| 2 .

4. An advertising placement plan generation system, including: A front-end module, a service processing module, and a data processing module; The front-end module is configured to send a request for generating a target advertising placement plan and display the target advertising placement plan; The business processing module is configured to receive the request for generating the target advertising placement plan, input training data to the data processing layer, and train and generate an adversarial network time series model according to the first historical advertising placement combination sequence data, preset parameters, and random vectors. Among them, training and generating an adversarial network time series model according to the first historical advertising placement combination sequence data, preset parameters, and random vectors includes: training the representation network in the adversarial network time series model according to the first historical advertising placement combination sequence data and the preset parameters: determining the first placement feature (S, X 1:T ) of the first historical advertising placement combination sequence data, and mapping the first placement feature (S, X 1:T ) from the original space to the latent space through the embedding network to obtain the corresponding first latent vector h S , h 1:T = e(s, x 1:t ), where S represents the vector space of static features, X 1:T represents the vector space with time series features, s ∈ S, x 1:T ∈ X 1:T , and e() is the embedding function representing the embedding network. The first latent vector h S , h 1:T = e(s, x 1:t ) is restored to the second placement feature in the original space r() is the restoration function representing the restoration network. The representation network includes the embedding network and the restoration network; training the adversarial network in the adversarial network time series model according to the random vector: The sequence generation network generates a second latent vector feature h S , h 1:T = g(z S , z 1:T ) in the latent space according to the random vector. z S represents a random vector sampled from a distribution, z t follows a random process, and g() is the function representing the sequence generation network. The sequence discriminator network discriminates the second latent vector feature h S , h 1:T = g(z S , z 1:T ) to obtain the sequence classification y S , y 1:T = d(h S , h 1:T ), h S = g S (z S ), h t = g X (h S , h t-1 , z t ), where d() is a function representing the sequence discriminant network, and g S : Z S →H S is a static feature generator, and g X : H S ×H S ×Z X →H X is a temporal feature recurrent generator, and H S 、H X represent the corresponding latent vector spaces of the feature spaces S and X 1:T . The adversarial network includes the sequence generation network and the sequence discriminant network. The supervision network is trained according to the representation network and the adversarial network. The representation network, the adversarial network, and the supervision network constitute the generative adversarial network temporal model; The data processing module is configured to receive the training data and the generative adversarial network time series model to obtain the target advertising placement plan and return the target advertising placement plan to the service processing module, wherein obtaining the target advertising placement plan includes: inputting the second historical advertising placement portfolio sequence data and target parameters into the generative adversarial network time series model to obtain a simulated target advertising placement portfolio: inputting the second historical advertising placement portfolio sequence data and synthetic data into the representation network, inputting the synthetic data of the second historical advertising placement portfolio sequence data into the adversarial network for joint training and iterating more than a preset threshold to generate simulated target advertising placement portfolio data; processing the simulated target advertising placement portfolio to determine the target advertising placement plan: inspecting the simulated target advertising placement portfolio to determine whether the simulated target advertising placement portfolio is real, and searching the simulated target advertising placement portfolio to output the optimal solution therein.

5. An electronic device, the electronic device including: One or more processors; A memory configured to store one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the advertising placement plan generation method according to any one of claims 1-3.

6. A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the advertising placement plan generation method according to any one of claims 1-3 is implemented.

Images