Coupon recommendation method based on anti-fact learning and multi-dimensional feature optimization
By using a method based on counterfactual learning and multi-dimensional feature optimization, and leveraging deep neural networks and generative adversarial networks to generate counterfactual samples, the problem of insufficient accuracy and coverage in traditional coupon recommendation methods is solved, achieving more accurate user behavior prediction and strategy adjustment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-10
AI Technical Summary
Traditional coupon recommendation methods struggle to capture the complex nonlinear relationships in user behavior and fail to fully utilize multi-dimensional features, resulting in low prediction accuracy and an inability to cover all user groups.
We employ a method based on counterfactual learning and multi-dimensional feature optimization. By acquiring users' historical consumption information, we construct a consumption amount sequence, analyze features using deep neural networks and dose-response networks, and combine generative adversarial networks to generate counterfactual samples, thereby dynamically adjusting the coupon distribution strategy.
It improves the accuracy and coverage of coupon recommendations, enables more precise prediction of user responses, and dynamically adjusts recommendation strategies to cover a wider user base.
Smart Images
Figure CN121836796A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of machine learning, in particular to a coupon recommendation method based on counterfactual learning and multi-dimensional feature optimization. BACKGROUND
[0002] With the rapid development of e-commerce platforms, personalized recommendation has become one of the key technologies to improve user experience and marketing effectiveness. Especially in marketing activities such as coupon distribution, how to accurately predict users' responses to different coupons based on their historical behaviors and characteristics, and dynamically adjust the recommendation strategy, has become a hot research topic. Traditional coupon recommendation methods mainly rely on rule-based algorithms or simple machine learning models. These methods often fail to capture the complex nonlinear relationships in user behavior, especially when faced with large-scale users and multi-dimensional features, the performance of the model is prone to bottlenecks.
[0003] Traditional methods usually model based on simple features of user behavior (such as purchase history, click behavior, etc.), and the recommendation algorithms used include collaborative filtering, content-based recommendation, decision trees, etc. Although these methods are effective in some cases, they have the following problems: first, traditional methods often ignore the complex nonlinear relationships in user behavior, resulting in low prediction accuracy, especially in complex situations; second, most traditional methods fail to fully utilize multi-dimensional features of users, especially ignoring the interaction effects and time series characteristics between features, resulting in an incomplete understanding of user preferences; third, traditional methods based on sample grouping often fail to make effective inferences when faced with rare events or unusual user behavior, resulting in a recommendation strategy that cannot cover all user groups. SUMMARY
[0004] The present application provides a coupon recommendation method based on counterfactual learning and multi-dimensional feature optimization to solve the problem of insufficient accuracy and coverage of existing recommendation methods. The method includes:
[0005] Obtaining historical consumption information of a user and preprocessing the historical consumption information, the historical consumption information including historical consumption amount;
[0006] Constructing a user historical consumption amount sequence based on the historical consumption amount of the user in multiple preset time periods, and extracting statistical features from the user historical consumption amount sequence;
[0007] Inputting the preprocessed historical consumption information into a trained deep neural network for analysis and processing to obtain consumption features of different dimensions;
[0008] The different-dimension consumption features and the coupon receiving information are input into the trained dose-response network for analysis and processing to obtain the counterfactual inference result, and a generative adversarial network is used to generate counterfactual samples;
[0009] The coupon issuing strategy is dynamically adjusted according to the counterfactual inference result and the counterfactual samples.
[0010] Optionally, the user historical consumption amount sequence is constructed according to historical consumption amounts of the user in a plurality of preset time periods, and the method comprises:
[0011] A first consumption amount sequence is constructed according to consumption amounts of the user in each day in the same month in the previous year and the current month;
[0012] A second consumption amount sequence is constructed according to consumption amounts of the user in each day in the previous N months of the current month;
[0013] The first consumption amount sequence and the second consumption amount sequence are spliced to obtain the user historical consumption amount sequence.
[0014] Optionally, the statistical features are extracted from the user historical consumption amount sequence, and the method comprises:
[0015] An average consumption amount of the user in the historical time period is determined according to consumption amounts of the user in each day in the historical time period and a total number of days;
[0016] A consumption volatility of the user in the historical time period is determined according to the consumption amounts of the user in each day in the historical time period, the total number of days and the average consumption amount.
[0017] Optionally, the preprocessed historical consumption information is input into the trained deep neural network for analysis and processing to obtain the different-dimension consumption features, and the method comprises:
[0018] The preprocessed historical consumption information is input into a deep neural network, and the deep neural network comprises an input layer, at least one hidden layer and an output layer, wherein a causal gating mechanism is introduced in the input layer, and the hidden layer adopts an attention mechanism;
[0019] The historical consumption information input is linearly transformed by the deep neural network to obtain a transformed feature matrix, and the transformed feature matrix is processed by a nonlinear activation function to obtain an output feature.
[0020] Optionally, the preprocessed historical consumption information is input into the trained deep neural network for analysis and processing, and the method further comprises:
[0021] The input historical consumption information is filtered by the causal gating mechanism to generate a gating feature matrix;
[0022] Determine an influence measure of different historical consumption information on user behavior results through a feature importance network;
[0023] Generate a query weight matrix, a key weight matrix and a value weight matrix dynamically according to the influence measure of the user behavior results through a learnable network;
[0024] Perform linear transformation on the gating feature matrix according to the generated query weight matrix, key weight matrix and value weight matrix to obtain a query matrix, a key matrix and a value matrix;
[0025] Calculate an attention score matrix according to the query matrix and the key matrix;
[0026] Normalize the attention score matrix through a Softmax function to generate an attention weight matrix;
[0027] Perform weighted summation on the value matrix using the attention weight matrix to obtain a final consumption feature.
[0028] Optionally, the analysis and processing of the consumption features and the coupon receiving information of different dimensions through the trained dose-response network to obtain the counterfactual inference result comprises:
[0029] Input the consumption features and the corresponding coupon receiving identifiers of the user into the trained dose-response network to predict the user behavior, and obtain a prediction result of the actual behavior of the user and a prediction result of the counterfactual behavior of the user.
[0030] Optionally, the generation of the counterfactual sample using the generative adversarial network comprises:
[0031] Sample a noise vector from a preset latent space, input the noise vector into a generator of the generative adversarial network, and generate a counterfactual sample through a generation function of the generator, wherein the counterfactual sample represents the potential consumption behavior of the user when the user does not receive the coupon;
[0032] Input the real sample and the generated counterfactual sample into a discriminator of the generative adversarial network, and output a probability score of the authenticity of the sample through a discrimination function of the discriminator.
[0033] The application provides a coupon recommendation method based on counterfactual learning and multi-dimensional feature optimization, which comprises the following steps: obtaining historical consumption information of a user, and preprocessing the historical consumption information, wherein the historical consumption information comprises historical consumption amount; constructing a user historical consumption amount sequence according to the historical consumption amount of the user in a plurality of preset time periods, and extracting statistical features from the user historical consumption amount sequence; inputting the preprocessed historical consumption information into a trained deep neural network for analysis and processing to obtain consumption features of different dimensions; inputting the consumption features of different dimensions and coupon receiving information into a trained dose-response network for analysis and processing to obtain a counterfactual inference result, and generating counterfactual samples using a generative adversarial network; determining a user behavior prediction result according to the counterfactual inference result and the counterfactual samples, and dynamically adjusting a coupon issuing strategy according to the user behavior prediction result, thereby improving the accuracy and effect of coupon recommendation. BRIEF DESCRIPTION OF DRAWINGS
[0034] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.
[0035] Figure 1 A flowchart of the coupon recommendation method based on counterfactual learning and multi-dimensional feature optimization provided by the embodiments of the application.
[0036] Through the above drawings, the specific embodiments of the application have been shown, and will be described in more detail hereinafter. These drawings and the written description are not intended to limit the scope of the inventive concept in any way, but to illustrate the inventive concept for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION
[0037] In order to make the objects, technical solutions and advantages of the application clearer, the technical solutions of the application will be described clearly and completely below with reference to the drawings in the application. Obviously, the described embodiments are only some of the embodiments of the application, but not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.
[0038] The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the application and the above drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein.
[0039] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean example, illustration, or description. Any embodiment or design scheme described as "exemplary" or "for example" in the present application should not be construed as more preferred or more advantageous than other embodiments or design schemes. Rather, the use of "exemplary" or "for example" is intended to present relevant concepts in a concrete manner.
[0040] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.
[0041] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.
[0042] Figure 1 The flowchart of the coupon recommendation method based on counterfactual learning and multi-dimensional feature optimization provided by the embodiments of the present application is shown. As shown in Figure 1 The coupon recommendation method based on counterfactual learning and multi-dimensional feature optimization provided by the embodiments of the present application includes:
[0043] S1: Obtain the historical consumption information of the user, and pre-process the historical consumption information, wherein the historical consumption information includes historical consumption amount.
[0044] Among them, the historical consumption information of the user may be, for example, the user's browsing of goods, the user's clicking of advertisements and coupons, the user's historical purchase behavior, the transaction amount, the purchase frequency, the user's search keywords, the search frequency, the user's received coupons, the coupon usage, the time of cancellation, and the user's social network information.
[0045] The pre-processing of the historical consumption information includes: filling the missing values of the numerical type features with the mean value, removing the samples with completely repeated field values, and standardizing the numerical type features.
[0046] S2: Construct a user historical consumption amount sequence according to the historical consumption amount of the user in a plurality of preset time periods, and extract statistical features from the user historical consumption amount sequence.
[0047] Specifically, the user historical consumption amount sequence is constructed according to historical consumption amounts of the user in a plurality of preset time periods, including: constructing a first consumption amount sequence according to consumption amounts of the user in each day in the same month in the last year and the current month; constructing a second consumption amount sequence according to consumption amounts of the user in each day in the last N months of the current month; and splicing the first consumption amount sequence and the second consumption amount sequence to obtain the user historical consumption amount sequence.
[0048] For example, the consumption amounts of the user in the same month last year and in the past three months are obtained, and the consumption amounts of the user at each time point in the past period of time are recorded as , wherein is a timestamp, indicating a day.
[0049] The consumption amount sequence of the same month last year is represented as:
[0050]
[0051] , wherein represents the consumption amount of the user on the th day of the same month last year.
[0052] The consumption amount sequence of the past three months is represented as:
[0053]
[0054] , wherein represents the consumption amount of the user on the th day of the past three months, represents the consumption amount of the user on the th day of the past two months, and represent the consumption amounts of the user on the th day of the past month, respectively.
[0055] The consumption amount sequence of the user in the same month last year and the consumption amount sequence of the user in the past three months are spliced, and the final consumption amount sequence is represented as:
[0056]
[0057] , wherein is a feature vector containing the consumption amounts of the same month last year and the past three months.
[0058] Specifically, statistical features are extracted from the user historical consumption amount sequence, including: determining the average consumption amount of the user in the historical time period according to the consumption amount and the total number of days of the user in each day in the historical time period; and determining the consumption volatility of the user in the historical time period according to the consumption amount, the total number of days and the average consumption amount of the user in each day in the historical time period.
[0059] Exemplarily, the average consumption amount satisfies the following formula:
[0060]
[0061] wherein, is valued as , for example, represents the consumption amount of the first day of the previous three months.
[0062] The consumption volatility satisfies the following formula:
[0063]
[0064] wherein, is valued as .
[0065] It can be understood that, by constructing the historical consumption amount sequence of the user, the consumption mode and periodic behavior of the user can be accurately captured, so as to better predict the reaction of the user under different coupon conditions; the extracted statistical features are used to enhance the performance of the model in processing the historical consumption data of the user.
[0066] S3: inputting the preprocessed historical consumption information into the trained deep neural network for analysis and processing to obtain consumption features of different dimensions.
[0067] wherein, the deep neural network comprises an input layer, at least one hidden layer and an output layer, and a causal gating mechanism is introduced in the input layer, and the hidden layer adopts an attention mechanism.
[0068] Exemplarily, after the historical consumption information is preprocessed, the input features of the deep neural network are obtained, in the deep neural network, the output of each layer is processed through a nonlinear activation function, and the input of the i-th layer is denoted as , and the output after linear transformation is , wherein is the weight matrix of the i-th layer, is a bias term. First, the output is transformed by a nonlinear activation function :
[0069]
[0070] wherein, uses a Tanh activation function, ;
[0071] The attention mechanism and the causal gating are used to capture the influence of different features on the user behavior, and the input feature is denoted as , The user feature input, the constructed user historical consumption amount sequence and the coupon information constitute the input, wherein is the number of features, is the feature dimension.
[0072] First, the input features are screened through a causal gating mechanism to obtain a gated feature matrix. Given the input features wherein The input features are sent into the gating weight generation network and are normalized:
[0073]
[0074] wherein a Softmax activation function is used; represents the gating weight generation network; , , represents the effective participation degree of the i-th feature to the user behavior result.
[0075] Then, the features are screened through gating, and the gating weight is applied to the matrix form:
[0076]
[0077] wherein, represents the gated feature matrix, at this time the feature is retained, the feature is strongly suppressed, and the intermediate value is that the feature is partially retained.
[0078] Subsequently, the importance of each input feature is evaluated through a feature importance network. The process is that the gated feature matrix is mapped to a unified semantic space through a mapping layer, and then the importance score of each feature is output through the feature importance network composed of multiple layers of perceptrons, wherein represents a convolutional neural network, represents a learnable network, and the present application adopts a multilayer perceptron.
[0079] Then, the weight matrix of the query, the key and the value is adaptively generated according to the estimation result, so that the weight matrix can be dynamically adjusted according to the input scene. The query weight matrix, the key weight matrix and the value weight matrix are represented by the following formula:
[0080] , ,
[0081] wherein, represents the influence degree of the feature to the behavior result, a learnable network with consistent network structure but different parameters .
[0082] Based on the above dynamically generated weight matrix, the gated features are linearly changed to obtain the query matrix, the key matrix and the value matrix:
[0083] , ,
[0084] Finally, the attention score is calculated:
[0085]
[0086] After the attention score is normalized by the softmax function, the output result is obtained by weighted summation of the value matrix using the attention weight:
[0087]
[0088]
[0089] wherein, represents the attention weight.
[0090] It can be understood that the deep learning network model combined with the nonlinear activation function can capture complex feature interactions and nonlinear relationships; the attention mechanism and the causal gating enable the model to dynamically identify which features have a greater impact on the coupon effect, thereby enhancing the understanding of user consumption behavior.
[0091] S4: input the consumption features and the coupon receiving information of different dimensions into the trained dose response network for analysis and processing to obtain the counterfactual inference result, and generate counterfactual samples using the generative adversarial network.
[0092] Specifically, the consumption features and the coupon receiving information of different dimensions are input into the trained dose response network for analysis and processing to obtain the counterfactual inference result, including: inputting the user's consumption features and the corresponding coupon receiving identifier into the trained dose response network for user behavior prediction to obtain the prediction result of the user's actual behavior and the prediction result of the user's counterfactual behavior.
[0093] For example, assume that a user has the consumption feature at time , and the counterfactual behavior is if the user does not receive a certain coupon, the goal of the dose response network (DRNet) is to predict the user's actual behavior and coupon receiving information In this process, behavior in different situations is predicted:
[0094] Actual behavior:
[0095]
[0096] in Indicates the user's time I received a coupon.
[0097] Counterfactual behavior:
[0098]
[0099] in Indicates the user's time I did not claim the coupon.
[0100] in, Indicates the user's time No coupon claimed. It is the prediction function of the DRNet model.
[0101] In an optional embodiment, the training process of the dose-response network includes: constructing a training dataset containing historical user characteristics, coupon intervention information, and corresponding actual consumption behavior labels; optimizing model parameters using adversarial training with the objective function of minimizing the prediction error between actual behavior and counterfactual behavior, so that the distribution difference between the generated counterfactual behavior and the real uninterrupted consumption data is less than a preset threshold.
[0102] Specifically, generating counterfactual samples using a generative adversarial network includes: sampling noise vectors from a predefined latent space, where the noise vectors follow a standard normal distribution; inputting the noise vectors into the generator of the generative adversarial network; and generating counterfactual samples through the generator's generation function. The counterfactual samples represent the user's potential consumption behavior when they have not received a coupon, and are expressed by the following formula.
[0103]
[0104] in, This indicates the counterfactual data output. This represents the generator function. It is noise sampled from the potential space.
[0105] Real samples and generated counterfactual samples are input into the discriminator of a generative adversarial network (GAN). The discriminator's discrimination function outputs a probability score indicating the authenticity of each sample. satisfy:
[0106] ;
[0107] .
[0108] In an optional embodiment, the training process of the generative adversarial network includes: taking the minimization of the distribution difference between the generated samples and the real non-intervened samples as the target of the generator, taking the maximization of the discrimination accuracy of the real samples and the generated samples as the target of the discriminator, and optimizing the model parameters of the generative adversarial network. The learning rates of the generator and the discriminator are dynamically adjusted according to the loss values of the generator and the discriminator. The loss function of the discriminator contains at least one of the authenticity loss, the distribution consistency loss and the causal constraint loss, wherein the authenticity loss adopts binary cross entropy loss to measure the classification accuracy of the discriminator on the real samples and the generated samples; the distribution consistency loss introduces Wasserstein distance or maximum mean difference to constrain the distribution similarity of the real samples and the generated samples; the causal constraint loss is based on the causal inference theory, and minimizes the conditional distribution difference of the real samples and the generated samples after balancing the covariates. When the loss of the discriminator is lower than a threshold, the learning rate of the discriminator is reduced to prevent overfitting; when the classification accuracy of the discriminator on the real samples and the generated samples does not improve for N consecutive rounds, the training is terminated and the current optimal model parameters are saved.
[0109] It can be understood that through the counterfactual learning framework based on DRNet, the potential behavior of the user under different coupon strategies can be estimated, and the generative adversarial network can generate counterfactual samples to make up for the defects of the traditional sample grouping method and improve the accuracy of counterfactual inference.
[0110] S5: determining a user behavior prediction result according to the counterfactual inference result and the counterfactual sample, and dynamically adjusting the coupon issuing strategy according to the user behavior prediction result.
[0111] Specifically, by combining the counterfactual sample generated by the generative adversarial network and the counterfactual inference result of the DRNet, a more accurate user behavior prediction result can be obtained.
[0112]
[0113] wherein, is the user behavior prediction result, is the counterfactual inference result of the DRNet, is the counterfactual sample generated by the generative adversarial network, is a weight coefficient for adjusting the importance of the traditional counterfactual inference and the generated counterfactual sample.
[0114] Further, the recommendation strategy is optimized according to the user behavior prediction result:
[0115] Add coupon data dimensions, build coupon face value candidate set, and pre-set multiple face value levels to provide users with more coupon choices:
[0116]
[0117] For example: 5 yuan, 10 yuan, 20 yuan, full reduction, etc. Then, the user's response to each type of coupon is dynamically output according to the user's consumption amount interval, so as to facilitate timely adjustment of the coupon amount;
[0118] Add coupon data dimensions, build coupon distribution time candidate window, and the distribution time can be divided into: immediate distribution, delayed distribution, and behavior triggered distribution (such as after browsing, clicking, and adding to cart). The best time for issuing coupons is dynamically determined by using the user's recent behavior state;
[0119] Combine the explicit reminder mode of high face value coupon pop-up window with the silent display mode of low face value, and link the face value and the distribution mode.
[0120] Subsequent online learning can be combined with real-time user feedback to continuously update the model, further improving the accuracy and effectiveness of personalized recommendation.
[0121] The method provided by the embodiment of the application comprises the following steps: obtaining historical consumption information of a user, and preprocessing the historical consumption information, wherein the historical consumption information comprises historical consumption amount; constructing a user historical consumption amount sequence according to the historical consumption amount of the user in a plurality of preset time periods, and extracting statistical features from the user historical consumption amount sequence; inputting the preprocessed historical consumption information into a trained deep neural network for analysis and processing to obtain consumption features of different dimensions; inputting the consumption features of different dimensions and coupon receiving information into a trained dose-response network for analysis and processing to obtain a counterfactual inference result, and generating counterfactual samples by using a generative adversarial network; determining a user behavior prediction result according to the counterfactual inference result and the counterfactual samples, and dynamically adjusting a coupon distribution strategy according to the user behavior prediction result, thereby improving the accuracy and effectiveness of coupon recommendation.
[0122] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The application is intended to cover any variations, uses or adaptive changes of this application following in general the principles of the application and including such as come within the scope of the application. The specification and examples are to be construed as merely illustrative of preferred embodiments of the application to cover all modifications and equivalents thereof falling within the scope of the application. The true scope and spirit of the application are indicated by the following claims.
[0123] It is to be understood that the application is not limited to the precise construction already described above and shown in the drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application should only be limited by the claims appended hereto.
Claims
1. A coupon recommendation method based on counterfactual learning and multi-dimensional feature optimization, characterized in that, The method comprises: obtaining historical consumption information of a user, and preprocessing the historical consumption information, wherein the historical consumption information comprises historical consumption amounts; constructing a user historical consumption amount sequence according to historical consumption amounts of the user in a plurality of preset time periods, and extracting statistical features from the user historical consumption amount sequence; inputting the preprocessed historical consumption information into a trained deep neural network for analysis and processing to obtain consumption features of different dimensions; inputting the consumption features of different dimensions and coupon receiving information into a trained dose-response network for analysis and processing to obtain counterfactual inference results, and generating counterfactual samples using a generative adversarial network; determining a user behavior prediction result according to the counterfactual inference results and the counterfactual samples, and dynamically adjusting a coupon issuing strategy according to the user behavior prediction result.
2. The method of claim 1, wherein, The method comprises: constructing a first consumption amount sequence according to consumption amounts of the user on each day in the same month of the current month in the previous year; constructing a second consumption amount sequence according to consumption amounts of the user on each day in the previous N months of the current month; splicing the first consumption amount sequence and the second consumption amount sequence to obtain the user historical consumption amount sequence.
3. The method of claim 2, wherein, The method comprises: determining an average consumption amount of the user in the historical time period according to the consumption amount of the user on each day in the historical time period and the total number of days; determining consumption volatility of the user in the historical time period according to the consumption amount of the user on each day in the historical time period, the total number of days, and the average consumption amount.
4. The method of claim 1, wherein, The method comprises: inputting the preprocessed historical consumption information into a deep neural network, wherein the deep neural network comprises an input layer, at least one hidden layer, and an output layer, wherein a causal gating mechanism is introduced in the input layer, and the hidden layer adopts an attention mechanism; performing linear transformation on the input historical consumption information through the deep neural network to obtain a transformed feature matrix, and processing the transformed feature matrix through a nonlinear activation function to obtain an output feature.
5. The method of claim 4, wherein, The method further comprises: filtering the input historical consumption information through the causal gating mechanism to generate a gated feature matrix; determining an influence measure of different historical consumption information on a user behavior result through a feature importance network; dynamically generating a query weight matrix, a key weight matrix, and a value weight matrix according to the influence measure of the user behavior result through a learnable network; performing linear transformation on the gated feature matrix according to the generated query weight matrix, the key weight matrix, and the value weight matrix to obtain a query matrix, a key matrix, and a value matrix; calculating an attention score matrix according to the query matrix and the key matrix; normalizing the attention score matrix through a Softmax function to generate an attention weight matrix; performing weighted summation on the value matrix using the attention weight matrix to obtain a final consumption feature.
6. The method of claim 1, wherein, The different dimension consumption characteristics and coupon receiving information are input into the trained dose response network for analysis and processing to obtain the counterfactual inference result, comprising: The consumption characteristics of the user and the corresponding coupon receiving identifier are input into the trained dose response network for user behavior prediction to obtain the prediction result of the actual behavior of the user and the prediction result of the counterfactual behavior of the user.
7. The method of claim 6, wherein, The counterfactual sample is generated by using the generative adversarial network, comprising: A noise vector is sampled from a preset latent space, the noise vector is input into a generator of the generative adversarial network, and a counterfactual sample is generated by a generation function of the generator, wherein the counterfactual sample represents potential consumption behavior of the user when the user does not receive the coupon; The real sample and the generated counterfactual sample are input into a discriminator of the generative adversarial network, and a probability score of the authenticity of the sample is output by a discrimination function of the discriminator.