Marketing advertisement click rate estimation method based on context feature interaction network

Abstract

The embodiment of the invention provides a marketing advertisement click rate estimation method based on a context feature interaction network, electronic equipment and a readable medium, and relates to the technical field of data processing, and the method comprises the steps: obtaining historical marketing advertisement click data; preprocessing the historical marketing advertisement click data; packaging the preprocessed historical marketing advertisement click data into a training sample; constructing a marketing advertisement click-through rate estimation model based on a context feature interaction network; performing click rate estimation on each training sample through a marketing advertisement click rate estimation model based on a context feature interaction network; according to the click rate estimation result of each training sample, training a marketing advertisement click rate estimation model based on the context feature interaction network; acquiring real-time marketing advertisement click data; and through the trained marketing advertisement click-through rate estimation model based on the context feature interaction network, performing click-through rate estimation on real-time marketing advertisement click-through data.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a marketing advertisement click-through rate prediction method based on a context feature interactive network, a marketing advertisement click-through rate prediction device based on a context feature interactive network, an electronic device, and a computer-readable medium. Background Art

[0002] The prediction of marketing advertisement click-through rate plays an irreplaceable role in the recommendation system. The prediction of marketing advertisement click-through rate aims to estimate the probability of users clicking on the recommended products on the web page. A more accurate click-through rate can not only bring higher revenue, but also greatly improve user satisfaction.

[0003] In order to improve the accuracy of click-through rate, researchers have proposed many methods to design more efficient interaction architectures. On the one hand, estimation methods based on factorization machines focus on learning low-order feature interactions to achieve lower time and space complexity. However, using only first-order and second-order feature interactions may limit the performance of the model and fail to achieve higher click-through rate accuracy. On the other hand, xDeepFM (Combining Explicit and Implicit Feature Interactions for Recommender Systems) and AutoInt (Automatic Feature Interaction Learning via Self-Attentive Neural Networks) use a two-stream structure design method to learn the cross-encoding of stream features through linear and nonlinear features. Although the two-stream MLP structure with different parameters and stacking depths can construct different nonlinear feature spaces to a certain extent, it is impossible to accurately obtain contextual interaction features, resulting in poor model click-through rate estimation performance. Summary of the invention

[0004] The embodiments of the present invention provide a marketing advertisement click-through rate prediction method based on a context feature interaction network, a marketing advertisement click-through rate prediction device based on a context feature interaction network, an electronic device, and a computer-readable storage medium, so as to solve the problem that the prior art uses a dual-stream MLP structure with different parameters and stacking depths and cannot accurately obtain context interaction features, resulting in poor model click-through rate prediction performance.

[0005] The embodiment of the present invention discloses a method for estimating click-through rate of marketing advertisements based on a context feature interactive network, comprising:

[0006] S1: Get historical marketing advertisement click data;

[0007] S2: Preprocessing the historical marketing advertisement click data;

[0008] S3: Package the preprocessed historical marketing advertisement click data into training samples;

[0009] S4: Construct a marketing advertisement click-through rate prediction model based on contextual feature interaction network;

[0010] S5: estimating the click rate of each training sample by using a marketing advertisement click rate prediction model based on a context feature interaction network;

[0011] S6: training a marketing advertisement click-through rate prediction model based on a context feature interaction network according to the click-through rate prediction results of each of the training samples;

[0012] S7: Obtain real-time marketing advertising click data;

[0013] S8: Estimating the click rate of the real-time marketing advertisement click data by using the trained marketing advertisement click rate prediction model based on the context feature interaction network.

[0014] The embodiment of the present invention discloses a marketing advertisement click rate estimation device based on a context feature interactive network, comprising:

[0015] The acquisition module is used to obtain historical marketing advertisement click data;

[0016] A preprocessing module, used for preprocessing the historical marketing advertisement click data;

[0017] A packaging module is used to package the pre-processed historical marketing advertisement click data into training samples;

[0018] A construction module for constructing a marketing advertisement click-through rate prediction model based on a contextual feature interaction network;

[0019] An estimation module, used to estimate the click rate of each training sample by using a marketing advertisement click rate estimation model based on a context feature interaction network;

[0020] A training module, used for training a marketing advertisement click rate prediction model based on a context feature interaction network according to the click rate prediction results of each training sample;

[0021] The acquisition module is also used to obtain real-time marketing advertisement click data;

[0022] The estimation module is also used to estimate the click rate of the real-time marketing advertisement click data through a trained marketing advertisement click rate estimation model based on a context feature interaction network.

[0023] The embodiment of the present invention further discloses an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0024] The memory is used to store computer programs;

[0025] The processor is used to implement the marketing advertisement click rate estimation method based on context feature interactive network as described in the embodiment of the present invention when executing the program stored in the memory.

[0026] The embodiment of the present invention further discloses one or more computer-readable media having instructions stored thereon, which, when executed by one or more processors, enable the processors to execute the method for estimating click-through rate of marketing advertisements based on a contextual feature interactive network as described in the embodiment of the present invention.

[0027] The embodiments of the present invention include the following advantages:

[0028] The present invention captures the relationship between contextual representation feature information and cross-features through a marketing advertisement click-through rate prediction model based on a contextual feature interaction network, significantly improves the interaction performance of complex features, and improves the accuracy of the model in estimating marketing advertisement click-through rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of the steps of a marketing advertisement click rate estimation method based on a context feature interactive network provided in an embodiment of the present invention;

[0030] Figure 2 is a schematic diagram of a context feature interaction network provided in an embodiment of the present invention;

[0031] Figure 3 is a schematic diagram of an attention network provided in an embodiment of the present invention;

[0032] Figure 4 is a schematic diagram of an attention-based complementary selection gating network provided in an embodiment of the present invention;

[0033] Figure 5 It is a block diagram of a marketing advertisement click rate estimation device based on a context feature interactive network provided in an embodiment of the present invention;

[0034] Figure 6 is a block diagram of an electronic device provided in an embodiment of the present invention;

[0035] Figure 7 is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Reference Figure 1 , showing a step flow chart of a method for estimating click-through rate of marketing advertisements based on a context feature interactive network provided in an embodiment of the present invention.

[0038] Reference Figure 2 , showing a schematic diagram of a context feature interaction network provided in an embodiment of the present invention.

[0039] Reference Figure 3 , showing a schematic diagram of an attention network provided in an embodiment of the present invention.

[0040] Reference Figure 4 , showing a schematic diagram of an attention-based complementary selection gating network provided in an embodiment of the present invention.

[0041] The marketing advertisement click rate estimation method based on context feature interaction network is applied to the cloud platform, which may specifically include the following steps:

[0042] S1: Get historical marketing ad click data.

[0043] Optionally, use Flink to read raw marketing ad click data in HDFS in real time.

[0044] Among them, Apache Flink is an open source distributed stream processing framework that supports both stream processing and batch processing. It has efficient and reliable distributed computing capabilities and is widely used in scenarios such as real-time data analysis and event-driven applications.

[0045] Among them, HDFS (Hadoop Distributed File System) is one of the core components of the Apache Hadoop project. It is a distributed file system that can efficiently store and process large-scale data.

[0046] S2: Preprocess historical marketing advertising click data.

[0047] In a possible implementation, S2 specifically includes sub-steps S201 to S203:

[0048] S201: Use Flink operators to convert historical marketing ad click data into ad click sequences in a schema structure.

[0049] Among them, Flink operators are the core components of Flink applications and are used to define the logic of stream data or batch data processing. By connecting data streams to various operators, users can implement complex data processing tasks. Flink operators provide flexible stream data processing capabilities and can complete data cleaning, conversion, and aggregation through various operations.

[0050] Furthermore, Flink operators include: data conversion operators, data flow operation operators, and state management operators.

[0051] The Schema structure is a structured description of data that defines the fields and attributes of the data. Schema is usually used to describe the name, type, and constraints of each field in a data table or data stream. The Schema structure helps define the data structure and fields to ensure data consistency during processing.

[0052] In practical applications, reading the original data in HDFS through Flink operators and parsing it into data sequences with Schema structures is a key step in implementing the advertising click-through rate prediction model.

[0053] In the present invention, after being structured, the data can be seamlessly integrated into the machine learning pipeline, reducing errors and improving the operability and scalability of the data.

[0054] S202: Eliminate abnormal data in the advertisement click sequence.

[0055] Specifically, first, the integrity of the data needs to be checked. If there are null values ​​or missing values ​​in key fields in the ad click records (such as user ID, ad ID, or click time), these incomplete records need to be directly removed because they cannot provide effective information for subsequent analysis and modeling.

[0056] Secondly, verify the format of the data to ensure that the values ​​of all fields meet the expected format requirements. For example, the click time should be a legal timestamp, and the advertising ID and user ID should be valid identifiers. If the format of the field is incorrect, the data should also be considered abnormal.

[0057] Duplicate data is a common problem and needs to be deduplicated. Duplicate ad click records may be caused by multiple writes to the system or network delays. You can check the combination of user ID, ad ID, and click time to determine whether it is a duplicate record and retain the unique record.

[0058] Then, you can perform a time range check based on the time range of the advertising campaign. If the click time exceeds the effective activity time period of the advertisement, it means that these records may be invalid or recorded by mistake and need to be filtered out.

[0059] If the frequency of user click behavior is abnormal, further inspection is required. For example, if the number of clicks on an advertisement by a single user in a unit time is far beyond the normal range, it may be crawler behavior or system abnormality. You can set a reasonable threshold to filter out these unreasonable records.

[0060] In addition, data distribution anomalies can also be processed. If certain features (such as click-through rate, ad impressions, etc.) deviate greatly from the overall distribution, statistical analysis (such as mean, standard deviation, quantile) can be used to set a reasonable range and remove outliers that exceed the range.

[0061] In the present invention, high-quality data helps to build a more robust advertisement click-through rate prediction model and improve the accuracy and stability of the model prediction.

[0062] S203: Use hot encoding to convert categorical features into different numbers.

[0063] Among them, one-hot encoding is a method of converting categorical features into numerical features, which is used to process non-numerical data that cannot be directly accepted by machine learning or deep learning models. By representing each category as a one-hot vector, one-hot encoding provides an efficient feature expression method for the model.

[0064] Optionally, obtain input data including user information attributes (such as province, age, etc.), characteristic attributes of products (such as color, price, etc.), and user click records, and pre-process the data to obtain a data set.

[0065] In the present invention, the hot-encoded numerical features enable the model to efficiently capture the impact of category features on click-through rate, thereby improving the expressive power of the features and the predictive performance of the model.

[0066] S3: Package the preprocessed historical marketing advertising click data as training samples.

[0067] Optionally, the original messy data is preprocessed and constructed into usable training sets, validation sets, and test sets.

[0068] Optionally, the data set is shuffled, 80% is randomly selected for model training, 10% is used for model validation, and 10% is used for model testing. Specifically, a random algorithm is used to select 80% of the data as the training set, and the remaining 20% ​​is randomly selected at a ratio of 1:1 and divided into a validation set and a test set.

[0069] S4: Construct a marketing advertising click-through rate prediction model based on contextual feature interaction network.

[0070] In a possible implementation, the architecture of a marketing advertisement click-through rate prediction model based on a contextual feature interaction network includes: an attention network and an attention complementary selection gating network.

[0071] Among them, the attention network is a mechanism that can dynamically assign importance weights to different features, emphasize key information and weaken irrelevant information. Its core idea is to imitate the human attention mechanism and focus on the most useful parts when processing data. It is used to capture the cross-relationships between all features, dynamically adjust the weights of features, and generate context-aware feature representations.

[0072] Among them, the attention complementary selection gating network is an enhanced network mechanism that uses the gating mechanism to dynamically screen original features and interactive features, select more important information and weaken redundant information, control the flow of information, and thus further improve the accuracy of prediction.

[0073] In the present invention, the combination of the attention network and the attention complementary selection gating network gives full play to the capabilities of contextual feature learning and dynamic information regulation, which can significantly improve the expression ability and prediction performance of complex features in the advertising click rate prediction model.

[0074] S5: The click rate of each training sample is estimated through the marketing advertisement click rate prediction model based on the context feature interaction network.

[0075] It should be noted that, first, the input high-dimensional sparse data is mapped to a low-dimensional dense feature space and converted into a feature vector; then, the feature information of different stages is generated through Hadamard product and matrix product operations to construct rich contextual feature information, and at the same time, the original features and complementary features are input into the attention network to dynamically adjust the feature weights to obtain more effective context-aware feature output results; finally, the final prediction result is obtained through conversion.

[0076] In a possible implementation, S5 specifically includes sub-steps S501 to S505:

[0077] S501: Convert the high-dimensional and sparse historical marketing advertisement click data into a low-dimensional and dense feature vector to obtain an embedding vector:

[0078]

[0079] Among them, X i represents the i-th input instance, x m Represents the input data of the mth domain, where m represents the size of the domain.

[0080] E i =[e1,e2,e3,...,e p ,...,e m ]

[0081] Among them, E i ∈R m×d represents the i-th embedding instance, e p ∈R d represents the embedding vector of the p-th domain, R represents the real number domain, and d represents the embedding dimension.

[0082] In the present invention, the embedding vector reduces the computational complexity by reducing the dimension while retaining the semantic information of the original features. The embedding vector can capture the potential semantic relationship between advertising data and user behavior features, laying the foundation for subsequent interaction and attention mechanism.

[0083] S502: At each interaction layer, use Hadamard product and matrix product in turn to determine the domain-level interaction features:

[0084]

[0085] Among them, H l represents the domain-level interaction feature, H l-1 represents the intermediate interaction feature, H 0 represents the initial input feature, ⊙ represents the Hadamard product, represents matrix product, W l , U l represents learnable parameters.

[0086] In the present invention, the cross-features between different domains (such as user feature domain, advertising feature domain, and context domain) are calculated through Hadamard product and matrix product. The high-order interaction information between features is extracted layer by layer to capture the potential nonlinear relationship between features. The multi-layer interaction design can gradually capture the feature associations from simple to complex and improve the model expressiveness.

[0087] S503: Calculate attention through the attention network:

[0088] α=Attention(Q,K,V)W p ∈R m×d

[0089] Q,K,V=FW Q ,FW K ,FW V ,

[0090]

[0091] Among them, α represents attention, Attention represents the attention function, Q represents the query matrix, K represents the key matrix, V represents the value matrix, F represents the input feature matrix, and W p represents the projection mapping matrix, W Qrepresents the transformation weight matrix of the query matrix, W K represents the transformation weight matrix of the key matrix, W V represents the conversion weight matrix of the value matrix, Softmax represents the Softmax activation function, T represents the matrix transpose operation, d s Represents the attention dimension.

[0092] In the present invention, the attention network expands the feature expression in different contexts by capturing the cross-feature relationship between all feature pairs, thereby realizing feature representation learning under contextual features. The feature weights are adjusted in different contexts to enhance the context perception ability of feature expression.

[0093] S504: According to the initial input features, the intermediate interaction features and the domain-level interaction features, the context features are determined by selecting the gating network through attention complementarity:

[0094] Γ=(H l-1 ⊙(1-σ(W l H l ))+H 0 ⊙σ(W l H l )+H l )α

[0095] Among them, Γ represents the context feature, σ represents the activation function, and W l represents the learnable weight parameter and α represents the attention.

[0096] It should be noted that three input channels can be set in the attention complementary selection gating network to input the initial input features, intermediate interaction features, and domain-level interaction features respectively.

[0097] In this paper, a novel gating mechanism is constructed through the attention complementary selection gating network to control the information flow and filter out more important feature representation information from the original and complementary features. The effective combination of original features and interactive features provides rich context perception capabilities and further improves the prediction accuracy.

[0098] S505: Estimating click rate based on context features:

[0099]

[0100] in, Represents the estimated click rate, Sigmoid represents the Sigmoid activation function, ε represents the weight matrix, and η represents the bias.

[0101] It should be noted that if Indicates that the corresponding ad instance is an ad that the user will click on.

[0102] In the present invention, a marketing advertisement click-through rate prediction model based on a context feature interaction network is used to capture the relationship between context representation feature information and cross-features, thereby significantly improving the interaction performance of complex features and improving the accuracy of the model in predicting marketing advertisement click-through rates.

[0103] S6: According to the click rate prediction results of each training sample, the marketing advertisement click rate prediction model based on the context feature interaction network is trained.

[0104] It should be noted that based on the initially set batch, the training set is divided into batches and input into the estimation model to obtain the model's predicted output, which is then differentially calculated with the actual result in the training set. The model is then iterated through backpropagation until the model's predicted result is infinitely close to the actual result.

[0105] In a possible implementation, S6 specifically includes sub-steps S601 and S602:

[0106] S601: Constructing the cross entropy loss function of the marketing advertisement click rate prediction model based on the context feature interaction network:

[0107]

[0108] Among them, Loss represents the cross entropy loss function, y i represents the actual value of the click rate of the i-th training sample, represents the estimated click rate of the ith training sample, log represents the logarithmic function with base 10, and n represents the total number of training samples.

[0109] S602: With the goal of minimizing the cross entropy loss function, a marketing advertisement click-through rate prediction model based on the context feature interaction network is trained.

[0110] In the present invention, the cross entropy loss function is applicable to binary classification tasks (such as ad clicks or no clicks), which can quantify the difference between the predicted value and the true value. When the model prediction is close to the true probability (ie, close to 0 or 1), the loss value is small; when the prediction deviates from the true value, the loss value increases rapidly. This feature enables the model to focus on reducing errors more quickly. By minimizing the cross entropy loss function, the model can continuously adjust its understanding of feature interactions and contextual weights, so that the final predicted click-through rate value is closer to the actual situation.

[0111] In the actual training process, in this embodiment, the training parameters include an embedding dimension of 50, a batch of 512, a number of training cycles of 8, an epoch size of 30, a learning rate of 0.002, and a cross entropy loss function is used as the objective function of the click-through rate prediction model.

[0112] Optionally, Logloss and AUC, two evaluation indicators widely adopted in the industry, can be used to evaluate the prediction effect of the model. The lower the Logloss value and the higher the AUC value, the better the prediction performance of the model.

[0113] S7: Obtain real-time marketing advertising click data.

[0114] S8: The click rate prediction model of marketing advertisements based on the contextual feature interaction network is trained to predict the click rate of real-time marketing advertisement click data.

[0115] In this invention, a new method is designed to process the complex features in marketing advertising data. By constructing a contextual feature interaction network and an attention complementary selection gating network, the problem of ignoring the different importance of each feature in different environments in the existing methods is solved, thereby significantly improving the accuracy of advertising click-through rate prediction. This innovative method can better learn and extract complex features in advertising data, thereby improving the accuracy of predicting advertising click-through rate.

[0116] The embodiments of the present invention include the following advantages:

[0117] The present invention captures the relationship between contextual representation feature information and cross-features through a marketing advertisement click-through rate prediction model based on a contextual feature interaction network, significantly improves the interaction performance of complex features, and improves the accuracy of the model in estimating marketing advertisement click-through rate.

[0118] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0119] In addition, refer to Figure 5 , showing a block diagram of a marketing advertisement click rate estimation device based on a context feature interactive network provided in an embodiment of the present invention.

[0120] The embodiment of the present invention discloses a marketing advertisement click rate prediction device 20 based on a context feature interactive network, comprising:

[0121] Acquisition module 201, used to acquire historical marketing advertisement click data;

[0122] A preprocessing module 202, used for preprocessing the historical marketing advertisement click data;

[0123] A packaging module 203 is used to package the pre-processed historical marketing advertisement click data into training samples;

[0124] A construction module 204 is used to construct a marketing advertisement click rate prediction model based on a context feature interaction network;

[0125] An estimation module 205, configured to estimate the click rate of each training sample by using a marketing advertisement click rate estimation model based on a context feature interaction network;

[0126] A training module 206, configured to train a marketing advertisement click rate prediction model based on a context feature interaction network according to the click rate prediction results of each training sample;

[0127] The acquisition module 201 is also used to acquire real-time marketing advertisement click data;

[0128] The estimation module 205 is further used to estimate the click rate of the real-time marketing advertisement click data by using the trained marketing advertisement click rate estimation model based on the context feature interaction network.

[0129] In a possible implementation manner, the preprocessing module 202 is specifically configured to:

[0130] Use Flink operators to convert the historical marketing advertisement click data into an advertisement click sequence in a schema structure;

[0131] Eliminate abnormal data in the advertisement click sequence;

[0132] Use hot encoding to convert categorical features into different numbers.

[0133] In a possible implementation, the architecture of the marketing advertisement click-through rate prediction model based on the context feature interaction network includes: an attention network and an attention complementary selection gating network.

[0134] In a possible implementation manner, the estimation module 205 is specifically configured to:

[0135] Convert the high-dimensional and sparse historical marketing advertising click data into a low-dimensional and dense feature vector to obtain the embedding vector:

[0136]

[0137] Among them, X i represents the i-th input instance, x m represents the input data of the mth domain, where m represents the size of the domain;

[0138] E i =[e1,e2,e3,...,ep ,...,e m ]

[0139] Among them, E i ∈R m×d represents the i-th embedding instance, e p ∈R d represents the embedding vector of the pth domain, R represents the real number domain, and d represents the embedding dimension;

[0140] At each interaction layer, the domain-level interaction features are determined using Hadamard product and matrix product in sequence:

[0141]

[0142] Among them, H l represents the domain-level interaction feature, H l-1 represents the intermediate interaction feature, H 0 represents the initial input feature, ⊙ represents the Hadamard product, represents matrix product, W l , U l represents learnable parameters;

[0143] Through the attention network, calculate the attention:

[0144] α=Attention(Q,K,V)W p ∈R m×d

[0145] Q,K,V=FW Q ,FW K ,FW V ,

[0146]

[0147] Among them, α represents attention, Attention represents the attention function, Q represents the query matrix, K represents the key matrix, V represents the value matrix, F represents the input feature matrix, and W p represents the projection mapping matrix, W Q represents the transformation weight matrix of the query matrix, W K represents the transformation weight matrix of the key matrix, W V represents the conversion weight matrix of the value matrix, Softmax represents the Softmax activation function, T represents the matrix transpose operation, d s represents the attention dimension;

[0148] Based on the initial input features, intermediate interaction features, and domain-level interaction features, the gating network is selected through attention complementarity to determine the contextual features:

[0149] Γ=(H l-1 ⊙(1-σ(W l H l ))+H 0 ⊙σ(W l H l )+H l )α

[0150] Among them, Γ represents the context feature, σ represents the activation function, and W l represents the learnable weight parameter, α represents attention;

[0151] According to the context features, the click rate is estimated:

[0152]

[0153] in, Represents the estimated click rate, Sigmoid represents the Sigmoid activation function, ε represents the weight matrix, and η represents the bias.

[0154] In a possible implementation, the training module 206 is specifically used to:

[0155] Construct the cross entropy loss function of the marketing advertisement click rate prediction model based on the context feature interaction network:

[0156]

[0157] Among them, Loss represents the cross entropy loss function, y i represents the actual value of the click rate of the i-th training sample, represents the estimated click rate of the i-th training sample, log represents the logarithmic function with base 10, and n represents the total number of training samples;

[0158] With the goal of minimizing the cross entropy loss function, a marketing advertisement click-through rate prediction model based on a contextual feature interaction network is trained.

[0159] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0160] The embodiments of the present invention include the following advantages:

[0161] The present invention introduces a multi-dimensional sliding window mechanism, which retains the key feature values ​​in the most recent time period. It captures the dynamic changes of key features before an error occurs through multiple sliding operations. At the same time, it uses a category-weighted roulette algorithm to screen the most relevant feature values ​​according to the weights, thereby dynamically capturing the change patterns of features over time and the contextual relationships, and improving the log error prediction capability.

[0162] In addition, refer to Figure 6 , shows a block diagram of an electronic device provided in an embodiment of the present invention. The embodiment of the present invention also provides an electronic device, including a processor 1301, a communication interface 1302, a memory 1303 and a communication bus 1304, wherein the processor 1301, the communication interface 1302, and the memory 1303 communicate with each other through the communication bus 1304,

[0163] Memory 1303, used for storing computer programs;

[0164] The processor 1301 is used to implement the following steps when executing the program stored in the memory 1303:

[0165] S1: Get historical marketing advertisement click data;

[0166] S2: Preprocessing the historical marketing advertisement click data;

[0167] S3: Package the preprocessed historical marketing advertisement click data into training samples;

[0168] S4: Construct a marketing advertisement click-through rate prediction model based on contextual feature interaction network;

[0169] S5: estimating the click rate of each training sample by using a marketing advertisement click rate prediction model based on a context feature interaction network;

[0170] S6: training a marketing advertisement click-through rate prediction model based on a context feature interaction network according to the click-through rate prediction results of each of the training samples;

[0171] S7: Obtain real-time marketing advertising click data;

[0172] S8: Estimating the click rate of the real-time marketing advertisement click data by using the trained marketing advertisement click rate prediction model based on the context feature interaction network.

[0173] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0174] The communication interface is used for communication between the above terminal and other devices.

[0175] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0176] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0177] Reference Figure 7 , showing a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. In another embodiment provided by the present invention, a computer-readable storage medium 1401 is also provided, in which instructions are stored, and when the instructions are executed on a computer, the computer executes the marketing advertisement click rate estimation method based on the context feature interactive network described in the above embodiment.

[0178] In another embodiment provided by the present invention, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute the marketing advertisement click rate prediction method based on context feature interactive network described in the above embodiment.

[0179] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0180] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0181] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0182] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A marketing advertisement click rate prediction method based on context feature interaction network, characterized in that: include: S1: Get historical marketing advertisement click data; S2: Preprocessing the historical marketing advertisement click data; S3: Package the preprocessed historical marketing advertisement click data into training samples; S4: Construct a marketing advertisement click-through rate prediction model based on contextual feature interaction network; S5: estimating the click rate of each training sample by using a marketing advertisement click rate prediction model based on a context feature interaction network; S6: training a marketing advertisement click-through rate prediction model based on a context feature interaction network according to the click-through rate prediction results of each of the training samples; S7: Obtain real-time marketing advertising click data; S8: Estimating the click rate of the real-time marketing advertisement click data by using the trained marketing advertisement click rate prediction model based on the context feature interaction network.

2. The method for estimating click-through rate of marketing advertisements based on contextual feature interaction network according to claim 1, characterized in that: The S2 specifically includes: S201: using a Flink operator to convert the historical marketing advertisement click data into an advertisement click sequence in a schema structure; S202: Eliminate abnormal data in the advertisement click sequence; S203: Use hot encoding to convert categorical features into different numbers.

3. The method for estimating click-through rate of marketing advertisements based on contextual feature interactive networks according to claim 1, characterized in that: The architecture of the marketing advertisement click rate prediction model based on context feature interaction network includes: an attention network and an attention complementary selection gating network.

4. The method for estimating click-through rate of marketing advertisements based on contextual feature interaction network according to claim 3, characterized in that: The S5 specifically includes: S501: Convert the high-dimensional and sparse historical marketing advertisement click data into a low-dimensional and dense feature vector to obtain an embedding vector: Among them, X i represents the i-th input instance, x m represents the input data of the mth domain, where m represents the size of the domain; AND i =[e1,e2,e3,…,e p ,…,and m ] Among them, E i ∈R m×d represents the i-th embedding instance, e p ∈R d represents the embedding vector of the pth domain, R represents the real number domain, and d represents the embedding dimension; S502: At each interaction layer, use Hadamard product and matrix product in turn to determine the domain-level interaction features: Among them, H l represents the domain-level interaction feature, H l-1 represents the intermediate interaction feature, H 0 represents the initial input feature, ⊙ represents the Hadamard product, represents matrix product, W l , U l represents learnable parameters; S503: Calculate attention through the attention network: α=Attention(Q,K,V)W p ∈R m×d Q,K,V=FW Q ,FW K ,FW V , Among them, α represents attention, Attention represents the attention function, Q represents the query matrix, K represents the key matrix, V represents the value matrix, F represents the input feature matrix, and W p represents the projection mapping matrix, W Q represents the transformation weight matrix of the query matrix, W K represents the transformation weight matrix of the key matrix, W V represents the conversion weight matrix of the value matrix, Softmax represents the Softmax activation function, T represents the matrix transpose operation, d s represents the attention dimension; S504: According to the initial input features, the intermediate interaction features and the domain-level interaction features, the context features are determined by selecting the gating network through attention complementarity: C=(H l-1 ⊙(1-σ(W l H l ))+H 0 ⊙σ(W l H l )+H l )a Among them, Γ represents the context feature, σ represents the activation function, and W l represents the learnable weight parameter, α represents attention; S505: Estimating the click rate based on the context features: in, Represents the estimated click rate, Sigmoid represents the Sigmoid activation function, ε represents the weight matrix, and η represents the bias.

5. The method for estimating click-through rate of marketing advertisements based on contextual feature interactive networks according to claim 1, characterized in that: The S6 specifically includes: S601: Constructing the cross entropy loss function of the marketing advertisement click rate prediction model based on the contextual feature interaction network: Among them, Loss represents the cross entropy loss function, y i represents the actual value of the click rate of the i-th training sample, represents the estimated click rate of the i-th training sample, log represents the logarithmic function with base 10, and n represents the total number of training samples; S602: With the goal of minimizing the cross entropy loss function, a marketing advertisement click-through rate prediction model based on a context feature interaction network is trained.

6. A marketing advertisement click rate prediction device based on context feature interactive network, characterized in that: include: The acquisition module is used to obtain historical marketing advertisement click data; A preprocessing module, used for preprocessing the historical marketing advertisement click data; A packaging module is used to package the pre-processed historical marketing advertisement click data into training samples; A construction module for constructing a marketing advertisement click-through rate prediction model based on a contextual feature interaction network; An estimation module, used to estimate the click rate of each training sample by using a marketing advertisement click rate estimation model based on a context feature interaction network; A training module, used for training a marketing advertisement click rate prediction model based on a context feature interaction network according to the click rate prediction results of each training sample; The acquisition module is also used to obtain real-time marketing advertisement click data; The estimation module is also used to estimate the click rate of the real-time marketing advertisement click data through a trained marketing advertisement click rate estimation model based on a context feature interaction network.

7. The marketing advertisement click rate prediction device based on context feature interactive network according to claim 6, characterized in that: The preprocessing module is specifically used for: Use Flink operators to convert the historical marketing advertisement click data into an advertisement click sequence in a schema structure; Eliminate abnormal data in the advertisement click sequence; Use hot encoding to convert categorical features into different numbers.

8. The marketing advertisement click rate prediction device based on context feature interactive network according to claim 6, characterized in that: The architecture of the marketing advertisement click rate prediction model based on context feature interaction network includes: an attention network and an attention complementary selection gating network.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the marketing advertisement click rate prediction method based on context feature interactive network as described in any one of claims 1 to 5 when executing the program stored in the memory.

10. A computer-readable medium, characterized in that Instructions are stored thereon, which, when executed by one or more processors, enable the processors to execute the marketing advertisement click rate prediction method based on a context feature interactive network as described in any one of claims 1-5.

Images