Advertisement pricing method and system
Through hybrid architecture neural networks, deep mining of the influencing factors of RTB advertising pricing has been solved, and the problem of low pricing accuracy of existing statistical models has been achieved, achieving more accurate advertising pricing.
Patent Information
- Application Number
- CN202411187662.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-12-20
AI Technical Summary
The existing RTB advertising pricing method based on statistical models cannot deeply explore the nonlinear relationship between influencing factors and pricing in complex data environments, resulting in low pricing accuracy and cannot meet the accuracy requirements of the ADX side.
Adopting hybrid architecture neural networks, including multi-layer perceptron networks, convolutional neural networks and long-term memory networks, train and verify historical advertising delivery data, build an advertising pricing model, and deeply explore the nonlinear relationship between influencing factors and pricing. , improve pricing accuracy.
Through deep learning technology, advertising pricing can be predicted more accurately, the accuracy of RTB advertising pricing can be improved, and the balance between the interests of the ADX side and the bidding success rate can be met.
Smart Images
Figure CN119168716A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of advertisement bidding, and more specifically, relates to an advertisement pricing method and system. Background Art
[0002] RTB advertising is a type of online advertising. RTB is the abbreviation of "Real Time Bidding", which means "real-time bidding". It is an emerging type of advertising that conducts real-time bidding based on each ad exposure.
[0003] Specifically, the bidding process of RTB advertising mainly includes four steps: the first step is that the SSP side (supply-side platform side) provides the user ID, advertising space information and minimum bid to the ADX side (advertising trading platform side); the second step is that the ADX side invites each DSP side to make a quotation; the third step is that each DSP side decides whether to bid and how much to bid by analyzing the value of the user; the fourth step is that the ADX side selects the highest advertising quotation from the advertising quotations of each DSP side, and determines the advertising quotation that is finally returned to the SSP side based on the highest advertising quotation of the DSP side and its own profit.
[0004] In the above-mentioned RTB advertising bidding process, the advertising price returned by the ADX end to the SSP end is one of the main factors that determine whether the RTB advertising can be successfully traded. If the advertising price returned by the ADX end to the SSP end is high, that is, the profit of the ADX end is set high, then the SSP end may refuse the transaction. If the advertising price returned by the ADX end to the SSP end is low, that is, the profit of the ADX end is set low, then the interests of the ADX end will be damaged. For this reason, the existing ADX end usually prices RTB advertising based on the RTB advertising pricing model to achieve a balance between its own interests and the success rate of RTB advertising bidding.
[0005] However, the existing RTB advertising pricing model is mainly a classic statistical model based on historical advertising data. Faced with an increasingly complex data environment, the RTB advertising pricing model based on statistical principles cannot deeply explore the nonlinear relationship between the various factors affecting RTB advertising pricing and the final pricing of RTB advertising, and to a certain extent, it cannot meet the ADX side's requirements for advertising pricing accuracy. Summary of the invention
[0006] The purpose of the present invention is to solve the problem of low pricing accuracy of the existing RTB advertisement pricing method based on statistical models.
[0007] To achieve the above objectives, the present invention provides an advertisement pricing method and system.
[0008] According to a first aspect of the present invention, there is provided an advertisement pricing method, the advertisement pricing method comprising the following steps:
[0009] A sample data set is obtained based on historical advertising data, where each piece of sample data in the sample data set includes a feature group and a label, where the feature group includes user click-through rate, advertising conversion rate, bidding success rate, and the highest advertising price on the DSP side, and the label is the advertising price, which is the advertising price returned by the ADX side to the SSP side;
[0010] The sample data set is used to train, verify and test the pre-established hybrid architecture neural network to obtain an advertising pricing model, wherein the advertising pricing model includes an input layer for receiving input features, a multi-layer perceptron network branch for fully connecting the input features to obtain high-order features, a convolutional neural network branch for performing convolution operations on the input features to obtain local features, a long short-term memory network branch for performing time series processing on the input features to obtain time dependencies, a fusion layer for splicing the output features of each network branch and fully connecting the spliced features, and an output layer for outputting the advertising pricing result;
[0011] The acquired feature group of the advertisement to be priced is input into the advertisement pricing model to obtain a corresponding advertisement pricing result, and the advertisement pricing result is returned to the SSP end.
[0012] Optionally, the acquiring of a sample data set based on historical advertisement delivery data includes:
[0013] Cleaning and normalizing the historical advertising data;
[0014] Feature extraction is performed on the normalized historical advertisement delivery data to obtain the sample data set.
[0015] Optionally, the multi-layer perceptron network branch is expressed as:
[0016]
[0017] In the above formula, h i MLP is the activation value of the i-th fully connected layer, W i MLP is the weight matrix of the i-th fully connected layer, X is the input feature group, b i MLP is the bias vector of the i-th fully connected layer, and RELU(·) is the activation function.
[0018] Optionally, the first two layers of the convolutional neural network branch are expressed as:
[0019]
[0020] In the above formula, h i CNN is the activation value of the i-th convolutional layer, W i CNN is the weight matrix of the i-th convolutional layer, b i CNN is the bias vector of the i-th convolutional layer, Conv(X,W) represents the convolution operation with the convolution kernel W on the input X, pi CNN is the pooling value of the i-th convolutional layer, and MaxPooling(h,k) represents the k*k maximum pooling operation on the input h.
[0021] Optionally, the first two layers of the LSTM network branch are expressed as:
[0022]
[0023] In the above formula, h 0 LSTM represents the hidden state vector at the initial time step, C 0 LSTM represents the memory cell state vector at the initial time step, h 1 LSTM represents the hidden state vector at the first time step, C 1 LSTM Represents the memory cell state vector at the first time step.
[0024] Optionally, the fusion layer is represented as:
[0025]
[0026] H fusion =ReLU(W fusion ·H concat +b fusion )
[0027] In the above formula, h k MLP is the output activation value of the multi-layer perceptron network branch, h m CNN is the output activation value of the convolutional neural network branch, h n LSTM is the output activation value of the LSTM network branch, H concat is the overall vector obtained by concatenating the output activation values of each network branch, W fusion is the weight matrix of the fusion layer, b fusion is the bias vector of the fusion layer, H fusion is the activation value of the fusion layer.
[0028] Optionally, the output layer is expressed as:
[0029] y=W out ·H fusion +b out
[0030] In the above formula, b out is the bias vector of the output layer, W out is the weight matrix of the output layer, and y is the advertising pricing result.
[0031] Optionally, the using the sample data set to train, verify and test the pre-established hybrid architecture neural network includes:
[0032] Dividing the sample data set into a training data set, a validation data set and a test data set;
[0033] Training the hybrid architecture neural network based on the training data set to obtain a hybrid architecture neural network to be verified;
[0034] Verifying the hybrid architecture neural network to be verified based on the verification data set, wherein the prediction error of the hybrid architecture neural network is evaluated based on a mean square error loss function, and the model parameters of the hybrid architecture neural network are adjusted using an Adam optimization algorithm to obtain the hybrid architecture neural network to be tested;
[0035] The hybrid architecture neural network to be tested is tested based on the test data set to obtain the advertising pricing model.
[0036] Optionally, the advertising pricing method further includes:
[0037] After returning the advertising pricing result each time, collect the actual delivery data of the corresponding advertisement;
[0038] According to a predetermined update cycle, the model parameters of the advertising pricing model are updated based on the collected actual delivery data.
[0039] According to a second aspect of the present invention, there is provided an advertisement pricing system, the advertisement pricing system comprising the following functional modules:
[0040] A sample data set acquisition module is used to acquire a sample data set based on historical advertising data. Each sample data in the sample data set includes a feature group and a label. The feature group includes user click-through rate, advertising conversion rate, bidding success rate and the highest advertising price on the DSP side. The label is the advertising price, which is the advertising price returned by the ADX side to the SSP side.
[0041] An advertising pricing model training module, used to train, test and verify a pre-established hybrid architecture neural network using the sample data set to obtain an advertising pricing model, wherein the advertising pricing model includes an input layer for receiving input features, a multi-layer perceptron network branch for fully connecting the input features to obtain high-order features, a convolutional neural network branch for performing convolution operations on the input features to obtain local features, a long short-term memory network branch for performing time series processing on the input features to obtain time dependencies, a fusion layer for splicing output features of each network branch and fully connecting the spliced features, and an output layer for outputting advertising pricing results;
[0042] The advertisement pricing result acquisition module is used to input the acquired feature group of the advertisement to be priced into the advertisement pricing model to obtain the corresponding advertisement pricing result, and return the advertisement pricing result to the SSP end.
[0043] The beneficial effects of the present invention are:
[0044] The advertising pricing method of the present invention first obtains a sample data set based on historical advertising delivery data; secondly, uses the sample data set to train, verify and test a pre-established hybrid architecture neural network to obtain an advertising pricing model; finally, inputs the acquired feature group of the advertisement to be priced into the advertising pricing model to obtain an advertising pricing result.
[0045] The advertising pricing method of the present invention is based on an advertising pricing model with a hybrid architecture neural network as the core to price RTB advertisements, and then returns the advertising pricing results to the SSP end. Compared with the existing RTB advertising pricing model based on statistical principles, the advertising pricing model based on a hybrid architecture neural network in the present invention can deeply explore the nonlinear relationship between various factors affecting RTB advertising pricing and the final pricing of RTB advertising, and its pricing of RTB advertising is relatively more accurate.
[0046] The advertising pricing system of the present invention and the above-mentioned advertising pricing method belong to a general inventive concept, and have at least the same beneficial effects as the above-mentioned advertising pricing method, and its beneficial effects are no longer repeated here.
[0047] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] The present invention may be better understood by referring to the following description taken in conjunction with the accompanying drawings, in which the same or similar reference numerals are used throughout the drawings to represent the same or similar components.
[0049] Figure 1A flowchart of an implementation method of an advertisement pricing method according to an embodiment of the present invention is shown;
[0050] Figure 2 A structural block diagram of an advertisement pricing system according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0051] In order to enable those skilled in the art to more fully understand the technical solution of the present invention, the exemplary embodiments of the present invention will be described in more comprehensive and detailed in conjunction with the accompanying drawings below. Obviously, the one or more embodiments of the present invention described below are only one or more of the specific ways in which the technical solution of the present invention can be implemented, and are not exhaustive. It should be understood that other ways belonging to a general inventive concept can be used to implement the technical solution of the present invention, and should not be limited by the exemplary embodiments described. Based on one or more embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work should fall within the scope of protection of the present invention.
[0052] Example: Figure 1 The following is a flowchart showing the implementation of the advertising pricing method according to an embodiment of the present invention. Figure 1 The advertising pricing method of the embodiment of the present invention comprises the following steps:
[0053] Step S100: Acquire a sample data set based on historical advertising data. Each piece of sample data in the sample data set includes a feature group and a label. The feature group includes user click-through rate, advertising conversion rate, bidding success rate, and the highest advertising price on the DSP side. The label is the advertising price, which is the advertising price returned by the ADX side to the SSP side.
[0054] Step S200: using a sample data set to train, verify and test a pre-established hybrid architecture neural network to obtain an advertising pricing model, the advertising pricing model includes an input layer for receiving input features, a multi-layer perceptron network branch for fully connecting the input features to obtain high-order features, a convolutional neural network branch for performing a convolution operation on the input features to obtain local features, a long short-term memory network branch for performing time series processing on the input features to obtain time dependencies, a fusion layer for splicing output features of each network branch and fully connecting the spliced features, and an output layer for outputting the advertising pricing result;
[0055] Step S300: input the acquired feature group of the advertisement to be priced into the advertisement pricing model to obtain a corresponding advertisement pricing result, and return the advertisement pricing result to the SSP end.
[0056] Furthermore, in step S100 of the embodiment of the present invention, obtaining a sample data set based on historical advertisement delivery data includes:
[0057] Clean and normalize historical advertising data;
[0058] Feature extraction is performed on the normalized historical advertising data to obtain a sample data set.
[0059] Specifically, in an embodiment of the present invention, historical advertising delivery data includes user click-through rate (Click-Through Rate, CTR), advertising conversion rate (Conversion Rate, CVR), bidding success rate and the highest advertising bid on the DSP side.
[0060] Specifically, in an embodiment of the present invention, a method of cleaning historical advertising delivery data is adopted to remove missing values and outliers in the historical advertising delivery data, thereby ensuring data quality; and a method of normalizing the cleaned historical advertising delivery data is adopted to make the data at the same level, so as to facilitate the training of the hybrid architecture neural network.
[0061] Specifically, in the embodiment of the present invention, a feature engineering method is used to extract features from normalized historical advertising data to select features that have a significant impact on advertising pricing.
[0062] Furthermore, in step S200 of the embodiment of the present invention, the multi-layer perceptron network branch is represented as:
[0063]
[0064] In the above formula, h i MLP is the activation value of the i-th fully connected layer, W i MLP is the weight matrix of the i-th fully connected layer, X is the input feature group, b i MLP is the bias vector of the i-th fully connected layer, and RELU(·) is the activation function.
[0065] Specifically, in the embodiment of the present invention, W i MLP Used to convert input features or the output of the previous layer into the output of the current layer; b i MLP Used to translate the output before the activation function to enhance the expressiveness of the model; h i MLP It is the output after being processed by the activation function, reflecting the nonlinear transformation of the input features by this layer; RELU(·) is used to introduce nonlinearity.
[0066] Furthermore, in step S200 of the embodiment of the present invention, the first two layers of the convolutional neural network branch are expressed as:
[0067]
[0068] In the above formula, h i CNN is the activation value of the i-th convolutional layer, W i CNN is the weight matrix of the i-th convolutional layer, b i CNN is the bias vector of the i-th convolutional layer, Conv(X,W) represents the convolution operation with the convolution kernel W on the input X, p i CNN is the pooling value of the i-th convolutional layer, and MaxPooling(h,k) represents the k*k maximum pooling operation on the input h.
[0069] Specifically, in the embodiment of the present invention, W i CNN Used to perform convolution operations on input features or the output of the previous layer to extract local features; b i CNN Used to adjust the result of the convolution operation; h i CNN is the output after convolution operation and activation function processing; p i CNN The maximum pooling is used to reduce the feature size to obtain the most important features.
[0070] Furthermore, in step S200 of the embodiment of the present invention, the first two layers of the long short-term memory network branch are expressed as:
[0071]
[0072] In the above formula, h 0 LSTM represents the hidden state vector at the initial time step, C 0 LSTM represents the memory cell state vector at the initial time step, h 1 LSTM represents the hidden state vector at the first time step, C 1 LSTM Represents the memory cell state vector at the first time step.
[0073] Furthermore, in step S200 of the embodiment of the present invention, the fusion layer is represented as:
[0074]
[0075] H fusion=ReLU(W fusion ·H concat +b fusion )
[0076] In the above formula, h k MLP is the output activation value of the multi-layer perceptron network branch, h m CNN is the output activation value of the convolutional neural network branch, h n LSTM is the output activation value of the LSTM network branch, H concat is the overall vector obtained by concatenating the output activation values of each network branch, W fusion is the weight matrix of the fusion layer, b fusion is the bias vector of the fusion layer, H fusion is the activation value of the fusion layer.
[0077] Specifically, in the embodiment of the present invention, h k MLP Contains the processing results of the multi-layer perceptron network branch on the input features, h m CNN Contains the processing results of the convolutional neural network branch on the input features, h n LSTM Contains the results of the long short-term memory network branch processing the input features.
[0078] Furthermore, in step S200 of the embodiment of the present invention, the output layer is represented as:
[0079] y=W out ·H fusion +b out
[0080] In the above formula, b out is the bias vector of the output layer, W out is the weight matrix of the output layer, and y is the advertising pricing result.
[0081] Furthermore, in step S200 of the embodiment of the present invention, using the sample data set to train, verify and test the pre-established hybrid architecture neural network includes:
[0082] Divide the sample data set into training data set, validation data set and test data set;
[0083] The hybrid architecture neural network is trained based on the training data set to obtain the hybrid architecture neural network to be verified;
[0084] The hybrid architecture neural network to be verified is verified based on the verification data set, wherein the prediction error of the hybrid architecture neural network is evaluated based on the mean square error loss function, and the model parameters of the hybrid architecture neural network are adjusted using the Adam optimization algorithm to obtain the hybrid architecture neural network to be tested;
[0085] The hybrid architecture neural network to be tested is tested based on the test data set to obtain the advertising pricing model.
[0086] Furthermore, the advertising pricing method of the embodiment of the present invention further includes the following steps:
[0087] After returning the advertising pricing result each time, collect the actual delivery data of the corresponding advertisement;
[0088] According to a predetermined update cycle, the model parameters of the advertising pricing model are updated based on the collected actual delivery data.
[0089] Specifically, in an embodiment of the present invention, a hybrid architecture neural network is used to construct an advertising pricing model. Among them, the parallel processing layer of the hybrid architecture neural network includes a multi-layer perceptron network branch, a convolutional neural network branch and a long short-term memory network branch; the multi-layer perceptron network branch is used to process high-order nonlinear features and capture complex relationships in the data; the convolutional neural network branch is used to extract local feature patterns and process spatial structural information in the data; the long short-term memory network branch is used to capture time series features and process time dependencies in the data. Through the combination of multi-layer perceptron network branches, convolutional neural network branches and long short-term memory network branches, a unique hybrid model architecture is formed, which improves the comprehensiveness and reliability of the advertising pricing model.
[0090] Accordingly, based on the advertisement pricing method of the embodiment of the present invention, the embodiment of the present invention also proposes an advertisement pricing system.
[0091] Figure 2 FIG. 1 shows a structural block diagram of an advertisement pricing system according to an embodiment of the present invention. Figure 2 The advertising pricing system of the embodiment of the present invention includes the following functional modules:
[0092] The sample data set acquisition module is used to acquire the sample data set based on the historical advertising data. Each sample data in the sample data set includes a feature group and a label. The feature group includes user click-through rate, advertising conversion rate, bidding success rate and the highest advertising price on the DSP side. The label is the advertising price, which is the advertising price returned by the ADX side to the SSP side.
[0093] An advertising pricing model training module is used to train, test and verify a pre-established hybrid architecture neural network using a sample data set to obtain an advertising pricing model. The advertising pricing model includes an input layer for receiving input features, a multi-layer perceptron network branch for fully connecting the input features to obtain high-order features, a convolutional neural network branch for performing convolution operations on the input features to obtain local features, a long short-term memory network branch for performing time series processing on the input features to obtain time dependencies, a fusion layer for splicing the output features of each network branch and fully connecting the spliced features, and an output layer for outputting the advertising pricing results;
[0094] The advertisement pricing result acquisition module is used to input the acquired feature group of the advertisement to be priced into the advertisement pricing model to obtain the corresponding advertisement pricing result, and return the advertisement pricing result to the SSP end.
[0095] Specifically, the advertising pricing system of the embodiment of the present invention adopts deep learning technology, which can extract complex high-order features from the data and increase the accuracy of the pricing model; combines the three models of MLP, CNN and LSTM to give full play to their respective advantages and improve the overall prediction ability of the model; uses mean square error (MSE) as the loss function, and continuously reduces the prediction error through back propagation and Adam optimizer to improve the accuracy of pricing.
[0096] Specifically, the advertising pricing system of the embodiment of the present invention, through the combination of data preprocessing, feature extraction and deep learning algorithms, makes full use of the historical data of advertising delivery, mines valuable information from massive data, and improves the utilization rate of data; comprehensively considers multiple factors such as user click-through rate (CTR), advertising conversion rate (CVR), bidding success rate, etc., conducts multi-level and multi-dimensional analysis, and makes the pricing strategy more scientific and reasonable.
[0097] Although one or more embodiments of the present invention are described above, it should be known to those skilled in the art that the present invention can be implemented in any other form without departing from its subject matter and scope. Therefore, the embodiments described above are illustrative rather than restrictive, and many modifications and substitutions are obvious to those skilled in the art without departing from the spirit and scope of the present invention as defined in the appended claims.
Claims
1. An advertising pricing method, characterized in that: include: A sample data set is obtained based on historical advertising data, where each piece of sample data in the sample data set includes a feature group and a label, where the feature group includes user click-through rate, advertising conversion rate, bidding success rate, and the highest advertising price on the DSP side, and the label is the advertising price, which is the advertising price returned by the ADX side to the SSP side; The sample data set is used to train, verify and test the pre-established hybrid architecture neural network to obtain an advertising pricing model, wherein the advertising pricing model includes an input layer for receiving input features, a multi-layer perceptron network branch for performing full-connection processing on the input features to obtain high-order features, a convolutional neural network branch for performing convolution operations on the input features to obtain local features, a long short-term memory network branch for performing time series processing on the input features to obtain time dependencies, a fusion layer for splicing output features of each network branch and performing full-connection fusion on the spliced features, and an output layer for outputting advertising pricing results; The acquired feature group of the advertisement to be priced is input into the advertisement pricing model to obtain a corresponding advertisement pricing result, and the advertisement pricing result is returned to the SSP end.
2. The advertising pricing method according to claim 1, characterized in that ,,Acquiring a sample data set based on historical advertising delivery data includes: Cleaning and normalizing the historical advertising data; Feature extraction is performed on the normalized historical advertisement delivery data to obtain the sample data set.
3. The advertising pricing method according to claim 1, characterized in that ,,The multi-layer perceptron network branch is expressed as: … In the above formula, h i MLP is the activation value of the i-th fully connected layer, W i MLP is the weight matrix of the i-th fully connected layer, X is the input feature group, b i MLP is the bias vector of the i-th fully connected layer, and RELU(·) is the activation function.
4. The advertising pricing method according to claim 3, characterized in that ,,The first two layers of the convolutional neural network branch are expressed as: In the above formula, h i CNN is the activation value of the i-th convolutional layer, W i CNN is the weight matrix of the i-th convolutional layer, b i CNN is the bias vector of the i-th convolutional layer, Conv(X,W) represents the convolution operation with the convolution kernel W on the input X, p i CNN is the pooling value of the i-th convolutional layer, and MaxPooling(h,k) represents the k*k maximum pooling operation on the input h.
5. The advertising pricing method according to claim 4, characterized in that ,,The first two layers of the long short-term memory network branch are expressed as: In the above formula, h0 LSTM represents the hidden state vector at the initial time step, C0 LSTM Represents the memory cell state vector at the initial time step, h1 LSTM represents the hidden state vector at the first time step, C1 LSTM Represents the memory cell state vector at the first time step.
6. The advertising pricing method according to claim 5, characterized in that , the fusion layer is expressed as: A fusion =ReLU(w fusion ·H concat +b fusion ) In the above formula, h k MLP is the output activation value of the multi-layer perceptron network branch, h m CNN is the output activation value of the convolutional neural network branch, h n LSTM is the output activation value of the LSTM network branch, H concat is the overall vector obtained by concatenating the output activation values of each network branch, W fusion is the weight matrix of the fusion layer, b fusion is the bias vector of the fusion layer, H fusion is the activation value of the fusion layer.
7. The advertising pricing method according to claim 6, characterized in that , the output layer is expressed as: y=W out ·H fusion +b out In the above formula, b out is the bias vector of the output layer, W out is the weight matrix of the output layer, and y is the advertising pricing result.
8. The advertising pricing method according to claim 7, characterized in that ,, the use of the sample data set to train, verify and test the pre-established hybrid architecture neural network includes: Dividing the sample data set into a training data set, a validation data set and a test data set; Training the hybrid architecture neural network based on the training data set to obtain a hybrid architecture neural network to be verified; Verifying the hybrid architecture neural network to be verified based on the verification data set, wherein the prediction error of the hybrid architecture neural network is evaluated based on a mean square error loss function, and the model parameters of the hybrid architecture neural network are adjusted using an Adam optimization algorithm to obtain the hybrid architecture neural network to be tested; The hybrid architecture neural network to be tested is tested based on the test data set to obtain the advertising pricing model.
9. The advertising pricing method according to claim 8, characterized in that ,, also includes: After returning the advertising pricing result each time, collect the actual delivery data of the corresponding advertisement; According to a predetermined update cycle, the model parameters of the advertising pricing model are updated based on the collected actual delivery data.
10. An advertising pricing system, characterized in that: include: A sample data set acquisition module is used to acquire a sample data set based on historical advertising data. Each sample data in the sample data set includes a feature group and a label. The feature group includes user click-through rate, advertising conversion rate, bidding success rate and the highest advertising price on the DSP side. The label is the advertising price, which is the advertising price returned by the ADX side to the SSP side. An advertising pricing model training module, used to train, test and verify a pre-established hybrid architecture neural network using the sample data set to obtain an advertising pricing model, wherein the advertising pricing model includes an input layer for receiving input features, a multi-layer perceptron network branch for fully connecting the input features to obtain high-order features, a convolutional neural network branch for performing convolution operations on the input features to obtain local features, a long short-term memory network branch for performing time series processing on the input features to obtain time dependencies, a fusion layer for splicing output features of each network branch and fully connecting the spliced features, and an output layer for outputting advertising pricing results; The advertisement pricing result acquisition module is used to input the acquired feature group of the advertisement to be priced into the advertisement pricing model to obtain the corresponding advertisement pricing result, and return the advertisement pricing result to the SSP end.