Consumption pattern recognition method and system

Abstract

The invention relates to the technical field of consumption analysis, in particular to a consumption pattern recognition method and system. The consumption pattern recognition system comprises a consumption behavior characteristic data acquisition module, a consumer group division module, a recommendation operation monitoring module and a causal effect value determination module. According to the method, the recommendation operation on the users is monitored, the prediction model and the intervention prediction model are constructed, and on the basis of the prediction model and the intervention prediction model, the causal effect value corresponding to each user is determined through an anti-fact inference technology, so that the sensitivity degree of each user to recommendation is determined, and then all consumer groups are re-divided; according to the method, the consumption behavior characteristic data in each consumer group is consistent, the sensitivity degree to recommendation is also consistent, it is ensured that each consumer group has similar key indexes such as the conversion rate when receiving the corresponding recommendation strategy, and the situation that the strategy is invalid due to the fact that the internal reaction difference of the consumer groups is too large is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of consumption analysis, in particular to a consumption pattern recognition method and system. BACKGROUND

[0002] With the rapid development of big data and artificial intelligence technology, personalized recommendation systems have been widely applied in e-commerce, finance, social platforms and other fields. In order to improve the recommendation effect, the platform usually needs to divide the groups according to the user's consumption behavior characteristics, that is, to divide the user consumption groups, so as to realize the personalized recommendation of different groups. However, the existing traditional clustering algorithm, especially the clustering method based on user behavior characteristics, has some challenges and limitations in consumption pattern recognition.

[0003] Traditional clustering methods, such as K-Means and Gaussian Mixture Model (GMM), mainly divide the groups through user behavior characteristics, such as purchase frequency, average order amount, category preference, etc. These methods usually assume that users in the same clustering group have high similarity in behavior. However, the actual situation is often more complex, and the user's recommendation response (such as conversion rate, repeat purchase rate, etc.) may not be completely consistent with the user's behavior characteristics.

[0004] A significant problem is that traditional clustering methods may divide users who are similar in behavior characteristics into the same group, but these users have large differences in response to recommendations. That is, although they show similarity in behavior characteristics such as purchase frequency, category preference, etc., their reactions to recommended content (such as whether to convert, whether to click, etc.) may differ significantly. This difference in causal effect is not effectively considered in traditional clustering, resulting in poor recommendation effect for the platform to these users. SUMMARY

[0005] The present application determines the sensitivity of each user to the recommendation by monitoring the recommendation operation on the user, constructing a prediction model and an intervention prediction model, and determining the causal effect value corresponding to each user through counterfactual inference technology based on the prediction model and the intervention prediction model. Then, all consumption groups are re-divided, so that the consumption behavior characteristic data in each consumption group is consistent, and the sensitivity to the recommendation is also consistent, ensuring that each consumption group has similar conversion rate and other key indicators when receiving the corresponding recommendation strategy, avoiding too large reaction difference within the consumption group, resulting in invalid strategy.

[0006] The present application provides a consumption pattern recognition method, comprising: Obtain consumption behavior feature data of all users, perform cluster analysis on the consumption behavior feature data of all users through a k-means clustering algorithm, and obtain a plurality of cluster clusters, each cluster cluster being recorded as a consumption group; Perform corresponding recommendation operations on the users in each consumption group, and continuously monitor the recommendation operations on the users within a preset time period to obtain recommendation feedback data, the recommendation feedback data including consumption behavior feature data, corresponding recommendation labels, and conversion probabilities; Determine a conversion prediction model and an intervention prediction model based on all recommendation feedback data, and determine a corresponding causal effect value for each user based on the prediction model and the intervention prediction model through counterfactual inference technology. Based on the corresponding causal effect value of each user, re-divide all consumption groups so that the consumption behavior feature data in each consumption group is consistent and the sensitivity to recommendations is also consistent.

[0007] As a preferred aspect, determining a conversion prediction model and an intervention prediction model based on all recommendation feedback data specifically includes the following steps: Obtain a pre-trained conversion prediction model and an intervention prediction model; Combine all recommendation feedback data into a prediction training set, train the pre-trained conversion prediction model through the prediction training set, and during the training, use the consumption behavior feature data and the corresponding recommendation labels in the recommendation feedback data as the input of the conversion prediction model, and use the conversion probability in the recommendation feedback data as the target output. Train the pre-trained intervention prediction model through the prediction training set, and during the training, use the consumption behavior feature data in the recommendation feedback data as the input of the intervention prediction model, and use the recommendation labels in the recommendation feedback data as the target output.

[0008] As a preferred aspect, based on the prediction model and the intervention prediction model, the corresponding causal effect value of each user is determined through counterfactual inference technology, specifically including the following steps: selecting any user, denoted as a target user, obtaining recommendation feedback data corresponding to the target user, denoted as target recommendation feedback data, traversing all the target recommendation feedback data, and performing the following operations for each target recommendation feedback data: feeding consumption behavior feature data in the target recommendation feedback data and a corresponding recommendation label into a conversion prediction model for processing to obtain a predicted conversion probability, and recording a difference between the conversion probability in the target recommendation feedback data and the predicted conversion probability as a first predicted residual; feeding the consumption behavior feature data in the target recommendation feedback data into an intervention prediction model for processing to obtain a predicted recommendation probability, and recording a difference between the recommendation label in the target recommendation feedback data and the predicted recommendation probability as a second predicted residual; calculating an average value of the first predicted residuals corresponding to all the target recommendation feedback data to obtain a conversion residual, calculating an average value of the second predicted residuals corresponding to all the target recommendation feedback data to obtain a recommendation residual, and using the conversion residual to regress the recommendation residual to obtain a causal effect value corresponding to each user.

[0009] As a preferred aspect, the conversion residual is used to regress the recommendation residual to obtain the causal effect value corresponding to each user, and the method specifically includes the following steps: The regression calculation is realized by the following formula: Y = a + bX + v, where Y is the conversion residual, X is the recommendation residual, a is a constant term, b is the causal effect value, and v is a regression error term.

[0010] As a preferred aspect, all consumption groups are re-divided based on the causal effect value corresponding to each user, and the method specifically includes the following steps: adding the causal effect value corresponding to the user at the end of the consumption behavior feature data corresponding to the user to form updated consumption behavior feature data, and then performing clustering analysis on the updated consumption behavior feature data of all users by using a k-means clustering algorithm to obtain a plurality of clustering clusters, each clustering cluster being recorded as a consumption group.

[0011] As a preferred aspect, the conversion prediction model and the intervention prediction model are both established based on a BP neural network.

[0012] The application also provides a consumption pattern recognition system, which includes: A consumption behavior feature data acquisition module is configured to obtain consumption behavior feature data of all users. A consumption group division module is configured to perform clustering analysis on the consumption behavior feature data of all users by using a k-means clustering algorithm to obtain a plurality of clustering clusters, each clustering cluster being recorded as a consumption group, and re-divide all consumption groups based on the causal effect value corresponding to each user. The recommended operation monitoring module is used for the user in each consumer group to perform the corresponding recommended operation, and continuously monitors the recommended operation to the user in a preset time period, obtains recommended feedback data, and the recommended feedback data includes consumer behavior characteristic data, a corresponding recommended label and a conversion probability. The causal effect value determination module is used for determining a conversion prediction model and an intervention prediction model based on all recommended feedback data, and determining the corresponding causal effect value of each user through the counterfactual inference technology based on the prediction model and the intervention prediction model.

[0013] The present application has the following advantages: The present application determines the sensitivity of each user to the recommendation by monitoring the recommended operation to the user, constructing the prediction model and the intervention prediction model, and determining the corresponding causal effect value of each user through the counterfactual inference technology based on the prediction model and the intervention prediction model, and then re-divides all consumer groups, so that the consumer behavior characteristic data in each consumer group is consistent, and the sensitivity to the recommendation is also consistent, ensuring that each consumer group has similar key indicators such as conversion rate when receiving the corresponding recommended strategy, avoiding that the internal reaction of the consumer group is too large, resulting in invalid strategy. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The structure diagram of the consumer mode recognition system used in the embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the person skilled in the art better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0016] Embodiment 1, a consumer mode recognition method, comprising: On the e-commerce platform, obtain the consumption behavior feature data of all users, the consumption behavior feature data including gender, age, geographic location, occupation, consumption frequency, order amount, category preference and active days, etc., wherein the consumption frequency refers to the number of times of completing purchase by the user within a certain period of time, for example, the number of times of completing purchase by the user within a month, which can be obtained through the order record of the user on the e-commerce platform, the order amount refers to the average consumption amount of the user when purchasing within a certain period of time, for example, the sum of all order amounts of the user within a month divided by the total number of purchase times within a month, the category preference indicates which types of goods the user tends to purchase on the e-commerce platform, which can be obtained by classifying the goods purchased by the user within a month, and the type of goods corresponding to the most purchase times is the category preference of the user, the category preference can be stored in the form of word embedding, and the active days refer to the number of days that the user actually logs in to the e-commerce platform within a certain period of time, these consumption behavior feature data can provide reference for active degree analysis, consumption capacity analysis and interest preference analysis of the user, so as to perform subsequent consumption mode recognition, and then the consumption behavior feature data of all users is analyzed by k-means clustering algorithm to obtain a plurality of clustering clusters, each clustering cluster is recorded as a consumption group, and each consumption group is defined as a consumption mode, the consumption group includes a plurality of consumption behavior feature data and corresponding user numbers, wherein the user number is the user number allocated when the user registers on the e-commerce platform; The specific use of the k-means clustering algorithm includes the following steps: Step S1: randomly select K consumption behavior feature data as the center from all the consumption behavior feature data of the users, the value of K can be determined by the elbow rule, wherein the elbow rule refers to calculating the total sum of squared errors (SSE) under different K values, and finding the inflection point (i.e. the "elbow") of the SSE to select the appropriate K value, since the elbow rule is a common prior art, it will not be described here; Step S2: add each consumption behavior feature data to the clustering cluster where the nearest center is located, wherein the distance refers to the Euclidean distance between the consumption behavior feature data and the center; Step S3: for each clustering cluster, calculate the average value of all consumption behavior feature data in the clustering cluster, and record it as a new center, to obtain K new centers; Step S4: repeat step S2 to step S3 until the center no longer changes significantly or the maximum number of iterations is reached, and output all clustering clusters; The corresponding recommendation operation is performed on the users in each consumer group, and the recommendation operation on the users in a preset time period is continuously monitored to obtain recommendation feedback data, which includes consumption behavior characteristic data, a corresponding recommendation label and a conversion probability. It should be noted that in the process of recommending to the user, a corresponding recommendation strategy can be set for each consumer group. Since the consumption behavior data stored in each consumer group is similar, the staff can set the recommendation strategy according to the corresponding consumption frequency, order amount, category preference and active days of the consumer group, for example, push the goods that meet the category preference, or push the recent promotion activities, and can determine which users to push in the background by the staff, or can determine which users to recommend by the recommendation algorithm built in the e-commerce platform. The recommendation label is generally represented by a Boolean value 1, and the conversion probability represents whether the user has a conversion behavior (such as purchase, collection, etc.) after receiving the recommendation. If a conversion behavior occurs, the conversion probability is 1, otherwise it is 0. The preset time period is generally one month, and the reconstruction of the consumer group is also one month apart. Based on all the recommendation feedback data, a conversion prediction model and an intervention prediction model are determined. The conversion prediction model is used to estimate the conversion probability of the user after receiving the recommendation, that is, to predict whether the recommendation will lead to the user's conversion (such as purchase, click, registration, etc.). The relationship between user features and conversion behavior can be established. The intervention prediction model is used to estimate whether the user will receive the recommendation, and the possibility of recommendation is determined based on user features. For example, a user is very active on the platform, has many purchase behaviors in the past 30 days, and often browses the goods on the platform, showing a clear purchase intention. The platform may consider him a "high-value user" and recommend related goods to him. Based on the prediction model and the intervention prediction model, the counterfactual inference technology is used to determine the causal effect value corresponding to each user, which represents whether the user will promote the user's conversion behavior after receiving the recommendation, that is, the sensitivity to the recommendation. Based on the causal effect value corresponding to each user, all consumer groups are re-divided so that the sensitivity to the recommendation in each consumer group is consistent with the consumption behavior characteristic data.

[0017] The present application determines the sensitivity of each user to the recommendation by monitoring the recommendation operation on the user, constructing a prediction model and an intervention prediction model, and determining the causal effect value corresponding to each user based on the prediction model and the intervention prediction model. Then all consumer groups are re-divided so that the sensitivity to the recommendation in each consumer group is consistent with the consumption behavior characteristic data, ensuring that each consumer group has similar conversion rates and other key indicators when receiving the corresponding recommendation strategy, avoiding too large differences in internal reactions of the consumer groups, resulting in invalid strategies.

[0018] The conversion prediction model and the intervention prediction model are determined based on all the recommendation feedback data, and specifically include the following steps: The conversion prediction model and the intervention prediction model are both established based on a BP neural network, and pre-trained conversion prediction models and intervention prediction models are obtained; All the recommendation feedback data are grouped into a prediction training set, and the pre-trained conversion prediction model is trained through the prediction training set. During the training, the consumption behavior feature data and the corresponding recommendation label in the recommendation feedback data are taken as the input of the conversion prediction model, and the conversion probability in the recommendation feedback data is taken as the target output. It is determined whether the training condition is met. The training condition is that the accuracy of the conversion prediction model meets the expectation. If the training condition is met, the construction of the conversion prediction model is completed. Otherwise, the pre-trained conversion prediction model is continuously trained through the prediction training set. The pre-trained intervention prediction model is trained through the prediction training set. During the training, the consumption behavior feature data in the recommendation feedback data is taken as the input of the intervention prediction model, and the recommendation label in the recommendation feedback data is taken as the target output. It is determined whether the training condition is met. The training condition is that the accuracy of the intervention prediction model meets the expectation. If the training condition is met, the construction of the intervention prediction model is completed. Otherwise, the pre-trained intervention prediction model is continuously trained through the prediction training set. It should be noted that the pre-trained conversion prediction model and the intervention prediction model refer to those trained through the recommendation feedback data in the historical records. The historical records here do not refer to one month, but a longer time dimension and a larger data volume. Based on the prediction model and the intervention prediction model, the causal effect value corresponding to each user is determined through counterfactual inference technology, and specifically includes the following steps: Select any user, denoted as a target user, obtain the recommendation feedback data corresponding to the target user, denoted as target recommendation feedback data, traverse all target recommendation feedback data, and perform the following operations for each target recommendation feedback data: input the consumption behavior feature data and the corresponding recommendation label in the target recommendation feedback data into the conversion prediction model for processing to obtain a predicted conversion probability, and record the difference between the conversion probability in the target recommendation feedback data and the predicted conversion probability as a first prediction residual, the first prediction residual representing the part that the conversion prediction model fails to accurately predict. If the conversion probability in the target recommendation feedback data is higher than the predicted conversion probability, it means that the conversion prediction model underestimates the conversion probability. Input the consumption behavior feature data in the target recommendation feedback data into the intervention prediction model for processing to obtain a predicted recommendation probability, and record the difference between the recommendation label in the target recommendation feedback data and the predicted recommendation probability as a second prediction residual, the second prediction residual representing the part that the intervention prediction model fails to accurately predict. If the recommendation label in the target recommendation feedback data is higher than the predicted recommendation probability, it means that the intervention prediction model underestimates the probability of receiving recommendations; calculate the average of the first prediction residuals corresponding to all target recommendation feedback data to obtain a conversion residual, calculate the average of the second prediction residuals corresponding to all target recommendation feedback data to obtain a recommendation residual, and use the conversion residual to regress the recommendation residual to obtain a causal effect value corresponding to each user.

[0019] The conversion residual is used to regress the recommendation residual to obtain a causal effect value corresponding to each user, specifically including the following steps: The regression calculation is realized by the following formula: Y = a + bX + v, where Y is the conversion residual, X is the recommendation residual, a is a constant term representing the baseline effect of residual regression, b is the causal effect value representing the causal influence of the recommendation operation on the conversion behavior, and v is the regression error term; The conversion residual reflects the difference between the user's actual conversion and the model's predicted conversion probability. This can help us understand whether the conversion behavior occurs due to other factors. The recommendation residual reflects the difference between whether the user receives the recommendation and the model's predicted recommendation probability. This can help us understand whether the recommendation works. Through the regression residual, we finally obtain b, which is the causal effect of the recommendation on the conversion, i.e., if the user receives the recommendation, the change in their conversion probability compared to the conversion probability when they do not receive the recommendation.

[0020] Based on the causal effect value corresponding to each user, all consumer groups are re-divided, specifically including the following steps: Add the user corresponding causal effect value at the end of the user corresponding consumption behavior characteristic data to form updated consumption behavior characteristic data, and then perform clustering analysis on the updated consumption behavior characteristic data of all users through a k-means clustering algorithm to obtain a plurality of clustering clusters, each clustering cluster being recorded as a consumption group.

[0021] Embodiment 2, a consumption pattern recognition system, see Figure 1 , comprising: The consumption behavior characteristic data acquisition module is configured to acquire consumption behavior characteristic data of all users on the e-commerce platform, the consumption behavior characteristic data including gender, age, geographic location, occupation, consumption frequency, order amount, category preference, and active days, etc., wherein the consumption frequency refers to the number of purchases completed by a user within a certain period of time, for example, the number of purchases completed by the user within a month, which can be acquired through order records of the user on the e-commerce platform, the order amount refers to the average consumption amount of the user when purchasing within a certain period of time, for example, the sum of all order amounts of the user within a month divided by the total number of purchases within a month, the category preference indicates which types of goods the user tends to purchase on the e-commerce platform, which can be determined by classifying the goods purchased by the user within a month, the type of goods corresponding to the highest number of purchases being the category preference of the user, the category preference can be stored in the form of word embedding, and the active days refer to the number of days that the user actually logs in to the e-commerce platform within a certain period of time, these consumption behavior characteristic data can provide a reference for active degree analysis, consumption ability analysis, and interest preference analysis of the user, so as to facilitate subsequent consumption pattern recognition; The consumption group division module is configured to perform clustering analysis on the consumption behavior characteristic data of all users through a k-means clustering algorithm to obtain a plurality of clustering clusters, each clustering cluster being recorded as a consumption group, and each consumption group being defined as a consumption pattern, the consumption group including a plurality of consumption behavior characteristic data and corresponding user numbers, the user numbers being assigned to the users when they register on the e-commerce platform, and each consumption group being re-divided based on the causal effect value corresponding to each user, so that the consumption behavior characteristic data in each consumption group is consistent, and the sensitivity to recommendations is also consistent. The recommendation operation monitoring module is configured to perform corresponding recommendation operations on users in each consumer group and continuously monitor the recommendation operations on the users in a preset time period to obtain recommendation feedback data, which includes consumer behavior characteristic data, corresponding recommendation labels and conversion probabilities. It should be noted that in the process of recommending to the users, a corresponding recommendation strategy is set for each consumer group. Since the consumer behavior data stored in each consumer group is similar, the staff can set the recommendation strategy according to the corresponding consumer frequency, order amount, category preference and active days of the consumer group, for example, push the goods that meet the category preference, or push the recent promotion activities, and can determine which users to push in the background by the staff, or can determine which users to recommend by the recommendation algorithm built in the e-commerce platform. The recommendation label is generally represented by Boolean value 1, and the conversion probability represents whether the user has a conversion behavior (such as purchase, collection, etc.) after receiving the recommendation. If the conversion behavior occurs, the conversion probability is 1, otherwise it is 0. The preset time period is generally one month, and the reconstruction of the consumer group is also one month apart. The causal effect value determination module is configured to determine a conversion prediction model and an intervention prediction model based on all the recommendation feedback data. The conversion prediction model is used to estimate the conversion probability of the user after receiving the recommendation, that is, to predict whether the recommendation will lead to the user's conversion (such as purchase, click, registration, etc.). The relationship between the user characteristics and the conversion behavior can be established. The intervention prediction model is used to estimate whether the user will receive the recommendation, and to determine the possibility of the recommendation based on the user characteristics. For example, a user is very active on the platform, has many purchase behaviors in the past 30 days, and often browses the goods on the platform, showing clear purchase intention. The platform may consider him a "high-value user" and recommend related goods to him based on these behavior characteristics. Based on the prediction model and the intervention prediction model, the counterfactual inference technology is used to determine the causal effect value corresponding to each user, which represents whether the user will promote the user's conversion behavior after receiving the recommendation, that is, the sensitivity to the recommendation.

[0022] It should be understood that those skilled in the art can make improvements or changes according to the above description, and all these improvements and changes should belong to the protection scope of the appended claims of the present application. The parts not described in detail in the specification belong to the prior art known to those skilled in the art.

Claims

1. A consumption pattern recognition method, characterized in that: include: Obtain consumption behavior feature data of all users, perform cluster analysis on the consumption behavior feature data of all users using the k-means clustering algorithm, and obtain several clusters, each of which is recorded as a consumer group; Execute corresponding recommended actions for users in each consumer group, and continuously monitor the recommended actions of users within a preset time period to obtain recommendation feedback data, which includes consumer behavior feature data, corresponding recommendation tags and conversion probability; Determine the conversion prediction model and intervention prediction model based on all recommendation feedback data. Based on the prediction model and intervention prediction model, determine the causal effect value corresponding to each user through counterfactual inference technology; All consumer groups are re-divided based on the causal effect value corresponding to each user, so that not only the consumption behavior characteristic data of each consumer group is consistent, but also the sensitivity to recommendations is consistent.

2. A consumption pattern recognition method according to claim 1, characterized in that: Determining the conversion prediction model and intervention prediction model based on all recommendation feedback data includes the following steps: Obtain pre-trained conversion prediction models and intervention prediction models; All recommendation feedback data is combined into a prediction training set, and the pre-trained conversion prediction model is trained using the prediction training set. During the training period, the consumer behavior feature data and corresponding recommendation labels in the recommendation feedback data are used as the input of the conversion prediction model, and the conversion probability in the recommendation feedback data is used as the target output; The pre-trained intervention prediction model is trained by predicting the training set. During the training, the consumer behavior feature data in the recommendation feedback data is used as the input of the intervention prediction model, and the recommendation label in the recommendation feedback data is used as the target output.

3. A consumption pattern recognition method according to claim 2, characterized in that: Based on the prediction model and intervention prediction model, the causal effect value corresponding to each user is determined through counterfactual inference technology, which includes the following steps: Select any user and record it as the target user. Obtain the recommendation feedback data corresponding to the target user and record it as the target recommendation feedback data. Traverse all target recommendation feedback data and perform the following operations for each target recommendation feedback data: send the consumption behavior feature data and the corresponding recommendation label in the target recommendation feedback data into the conversion prediction model for processing to obtain the predicted conversion probability, and record the difference between the conversion probability in the target recommendation feedback data and the predicted conversion probability as the first prediction residual. Send the consumption behavior feature data in the target recommendation feedback data into the intervention prediction model for processing to obtain the predicted recommendation probability, and record the difference between the recommendation label in the target recommendation feedback data and the predicted recommendation probability as the second prediction residual; calculate the average of the first prediction residuals corresponding to all target recommendation feedback data to obtain the conversion residual, calculate the average of the second prediction residuals corresponding to all target recommendation feedback data to obtain the recommendation residual, use the conversion residual to regress the recommendation residual to obtain the causal effect value corresponding to each user.

4. A consumption pattern recognition method according to claim 3, characterized in that: Use the transformation residuals to regress the recommendation residuals to obtain the causal effect value corresponding to each user. The specific steps include the following: The regression calculation is achieved through the following formula: Y=α+βX+v, where Y is the transformation residual, X is the recommended residual, α is the constant term, β is the causal effect value, and v is the regression error term.

5. A consumption pattern recognition method according to claim 4, characterized in that: All consumer groups are re-divided according to the causal effect value corresponding to each user, which specifically includes the following steps: adding the causal effect value corresponding to the user to the end of the consumption behavior characteristic data corresponding to the user to form the updated consumption behavior characteristic data, and then clustering analysis is performed on the updated consumption behavior characteristic data of all users through the k-means clustering algorithm to obtain several clusters, and each cluster is recorded as a consumer group.

6. A consumption pattern recognition method according to claim 5, characterized in that: Both the conversion prediction model and the intervention prediction model are established based on BP neural network.

7. A consumption pattern recognition system, characterized in that: The system applies a consumption pattern recognition method according to any one of claims 1 to 6, including: The consumption behavior characteristic data acquisition module is used to obtain the consumption behavior characteristic data of all users; The consumer group segmentation module is used to perform cluster analysis on the consumption behavior characteristic data of all users using the k-means clustering algorithm to obtain several clusters, each of which is recorded as a consumer group; and all consumer groups are re-divided based on the causal effect value corresponding to each user; The recommended action monitoring module is used to monitor the recommended actions taken by users in each consumer group and continuously monitor the recommended actions taken by users within a preset time period to obtain recommendation feedback data, which includes consumer behavior feature data, corresponding recommendation tags, and conversion probability. The causal effect value determination module is used to determine the conversion prediction model and intervention prediction model based on all recommendation feedback data. On the basis of the prediction model and intervention prediction model, the causal effect value corresponding to each user is determined through counterfactual inference technology.

Images