A digital marketing accurate pushing optimization method based on user behavior portrait

Abstract

This invention discloses a method for optimizing precise push notifications in digital marketing based on user behavior profiles, belonging to the field of digital marketing technology. The invention includes three steps: Step 1, user behavior profile construction and updating; Step 2, profile tag matching and push execution; and Step 3, push effect evaluation and rule optimization. By constructing dynamic user preference features and separately handling active closure and unsubscription behaviors, it overcomes the limitations of traditional methods, such as the contradiction between attenuation and trend characterization and the dilution of negative feedback. By using users who have not received push notifications as a control group for incremental evaluation and by assigning graded weights to different negative feedback behaviors, it breaks through the bottleneck of coarse optimization direction determination and difficulty in distinguishing push increments from conversions. Parameter optimization is performed based on the combination of conversion increments and negative feedback rates, balancing positive conversion improvement and user disturbance control. Closed-loop self-optimization is achieved through negative feedback-driven hierarchical mapping rule updates, improving user experience and marketing resource utilization.

Description

Technical Field

[0001] This invention relates to the field of digital marketing technology, specifically to a method for optimizing precise digital marketing push based on user behavior profiles. Background Technology

[0002] With the deep integration of internet technology and digital marketing methods, personalized and precise push notifications based on user behavior data have become a core operational approach in e-commerce, content platforms, and other fields. Existing technologies typically collect users' historical browsing, click, and purchase behavior data to build user profiles, and then match and push content based on these profiles. In profile building, to reflect the time-sensitive changes in user interests, a time decay is generally applied to historical behaviors. However, if a global decay is applied to each behavior individually first, and then the decayed behaviors are aggregated into different time windows, the weight of behaviors within early windows approaches zero, resulting in a loss of differentiation between windows and failing to accurately reflect the evolution of user preferences. Simultaneously, negative feedback behaviors such as unsubscribing from push notifications and closing content are clear and long-term stable rejection signals. Applying the same time decay to these signals causes content types that were explicitly rejected to re-enter the candidate push range over time, impairing push accuracy and user experience.

[0003] In terms of push notification execution and performance optimization, existing technologies typically adjust parameters based on metrics such as conversion rate and click-through rate after the push notification. However, conversion rate calculations often lack a reasonable benchmark, making it difficult to distinguish between the incremental effect brought by the push notification itself and the difference between natural user conversion, which can easily lead to deviations in optimization direction. For negative feedback behaviors such as user closure and unsubscription triggered by push notifications, existing evaluation methods often simply accumulate or treat them equally, failing to effectively differentiate between different levels of aversion, resulting in imprecise optimization decisions. Furthermore, common parameter adjustment strategies often apply uniform rules to all users, making it difficult to achieve a dynamic balance between improving conversion results and controlling user disturbance. Summary of the Invention

[0004] The purpose of this invention is to provide a digital marketing precision push optimization method based on user behavior profiles, which solves the problems existing in the background technology.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a digital marketing precision push optimization method based on user behavior profiles, including: step one, user behavior profile construction and updating, step two, profile tag matching and push execution, and step three, push effect evaluation and rule optimization.

[0006] Step 1: User Behavior Profile Construction and Update. Collect static attribute data and dynamic behavior data from the entire digital marketing process. Based on the dynamic behavior data, divide the time window into multiple consecutive time windows according to the time of behavior occurrence. Aggregate the dynamic behavior data in each window to obtain the original behavior feature representation. Apply time decay weight to each original behavior feature representation according to the time distance between each time window and the current time to form the user dynamic preference feature. Construct a comprehensive user profile based on the static attribute data and the user dynamic preference feature, and classify users into corresponding hierarchical labels. As time goes by, slide the time window to dynamically update the comprehensive user profile and hierarchical labels. Step 2: Profile Tag Matching and Push Execution. Based on user profiles and hierarchical tags, establish mapping rules between hierarchical tags and push content. Match corresponding push content to users according to the mapping rules. Set push rules based on user behavior characteristics. Filter users according to the push rules and execute pushes. Record push response data. Step 3: Evaluation of push effect and optimization of rules. Based on the recorded push response data, calculate the conversion increment and negative feedback rate. Adjust the time window length and hierarchical tag division according to the changes in conversion increment and negative feedback rate, and optimize the mapping rules and push rules.

[0007] The beneficial effects of this invention are as follows: (1) The present invention constructs a profile by first aggregating behaviors according to time windows and then applying time decay to the windows, so that the intensity of behaviors in different time windows can be fully preserved and the trend of temporal change is not covered by decay. At the same time, negative feedback behaviors such as actively unsubscribing and closing are processed separately, and their negative weights do not decay over time, so that users' long-term and stable content rejection preferences can be permanently remembered, avoiding the wrong re-pushing of content types that have been clearly rejected due to the passage of time, and improving the overall accuracy of digital marketing push.

[0008] (2) This method calculates the conversion increment by using users in the same stratum with similar profiles who have not been pushed to the control group in the effect evaluation stage, so as to truly reflect the marginal effect brought by the push itself. The closing and unsubscribing behaviors are included in the negative feedback rate with different weights, which can distinguish different degrees of aversion. Based on the positive and negative combination of conversion increment and negative feedback rate, four optimization branches are divided. The key parameters such as time window length, stratified tag division, push frequency limit, and mapping rules are adjusted in a targeted manner to achieve a dynamic balance between improving conversion effect and controlling user disturbance. This enables the system to continuously adapt to changes in marketing scenarios and user feedback, and achieve continuous self-optimization of accurate push. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Reference Figure 1 As shown, this invention provides a method for optimizing precise digital marketing push based on user behavior profiles, including: step one, user behavior profile construction and updating; step two, profile tag matching and push execution; and step three, push effect evaluation and rule optimization.

[0013] Step 1: User Behavior Profile Construction and Update. Collect static attribute data and dynamic behavior data from the entire digital marketing process. Based on the dynamic behavior data, divide the time window into multiple consecutive time windows according to the time of behavior occurrence. Aggregate the dynamic behavior data in each window to obtain the original behavior feature representation. Apply time decay weight to each original behavior feature representation according to the time distance between each time window and the current time to form the user dynamic preference feature. Construct a comprehensive user profile based on the static attribute data and the user dynamic preference feature, and classify users into corresponding hierarchical labels. As time goes by, slide the time window to dynamically update the comprehensive user profile and hierarchical labels. In the above embodiments, the static attribute data includes identity attribute data and long-term preference setting data; The dynamic behavioral data includes page view and click data, search query data, dwell and browsing depth data, add-to-cart and order data, push notifications, and close and unsubscribe data. The dynamic behavior data includes the time when the behavior occurred and the content attribute tags of the behavior object. The content attribute tags include content type tags and timeliness indicators.

[0014] In the above embodiments, the method of collecting static attribute data and dynamic behavior data of the entire digital marketing process of users, dividing the dynamic behavior data into multiple consecutive time windows according to the time of behavior occurrence, aggregating the dynamic behavior data in each window to obtain the original behavior feature representation, and applying a time decay weight to each original behavior feature representation according to the time distance between each time window and the current time to form the user dynamic preference feature, is as follows: The system performs missing value imputation and duplicate value removal on dynamic behavior data, normalizes continuous numerical indicators, and performs one-hot encoding on discrete content attribute tags to form standardized dynamic behavior features. A fixed-length time window is set, and each standardized dynamic behavior feature is assigned to a corresponding time window based on the occurrence time of the behavior. Within each time window, a positive initial weight is set according to the behavior type of each behavior feature, while the initial weights for closing and unsubscribing behaviors are set to negative values. The absolute value of the negative weight for unsubscribing behavior is greater than that for closing behavior. The negative weights are directly accumulated into the corresponding associated content attribute tags and are not decayed with the time window. Within each time window, the content attribute tags and their corresponding positive initial weights are integrated to form the window's original behavior feature vector, which has the same dimension as the content attribute tags. Based on the time distance between each time window and the current time, the time decay weight corresponding to each window is calculated using an exponential decay function. The decay weight value ranges from [value missing]. The closer the window is to the current time, the greater the decay weight. The decay weight of the current window is 1. The original behavioral feature vector of each window is multiplied by the corresponding decay weight. The feature vectors of all windows after decay processing are summed and the cumulative negative weight vectors over the whole period are added to obtain the user dynamic preference features.

[0015] It should be noted that static attribute data and dynamic behavior data are used to distinguish different time-sensitive attributes of user basic characteristics and behavioral preference characteristics. Static attribute data, such as identity information and long-term preference settings, are relatively stable, while dynamic behavior data is continuously updated with user behavior and can reflect the migration process of user interests. The content attribute tags in dynamic behavior data are further divided into content type tags and timeliness indicators. Content type tags are used to describe the content category of the behavior object, while timeliness indicators are used to mark whether the pushed content is time-sensitive.

[0016] It should be noted that: missing value imputation uses the average of similar behavioral features or preset default values; duplicate value removal is performed by comparing the consistency between the behavior occurrence time and the content attribute tags; continuous numerical index normalization maps continuous index values ​​to a preset unified numerical range; discrete content attribute tag one-hot encoding is performed by establishing a binary dimension for each content attribute tag, setting the corresponding dimension as a valid identifier when the tag is matched, and setting it as an invalid identifier when it is not matched; behavioral features are divided and aggregated according to fixed time windows to ensure that the behavioral intensity within each window is fully preserved; initial weights are set according to behavior type, with higher weights given to deep interactive behaviors such as purchasing and adding to cart, and lower weights given to shallow behaviors such as browsing and clicking; the initial weights of closing and unsubscribing behaviors are set to negative values ​​and they are excluded from the time decay rules; positive behaviors decay over time in line with the objective law of user interest migration, but closing and unsubscribing are long-term stable content rejection signals actively expressed by users, and not applying time decay can ensure that such negative preferences are permanently remembered by the profiling system, preventing rejected content from re-entering the candidate push range.

[0017] In the above embodiments, the method for constructing a comprehensive user profile based on static attribute data and dynamic user preference features, dividing users into corresponding hierarchical labels, and dynamically updating the comprehensive user profile and hierarchical labels over time by sliding a time window is as follows: Static attribute data is encoded to obtain static feature vectors. These static feature vectors are then concatenated with user dynamic preference features. The concatenated features are integrated to form a comprehensive user profile vector. All content attribute labels with the maximum weight in the user dynamic preference features are extracted as the user's dominant preference labels. Based on the dominant preference labels and long-term preference settings in the static attribute data, users are assigned to corresponding hierarchical labels. As time progresses, a sliding time window is used to incorporate dynamic behavior data from the new window and remove dynamic behavior data from the oldest window. The user dynamic preference features are then recalculated, and the comprehensive user profile vector and hierarchical labels are updated.

[0018] It should be noted that the encoding process converts the discrete category fields in the identity attribute data into binary feature dimensions, maps the preference labels in the long-term preference setting data to a label space that is consistent with the content attribute labels, forming a static feature vector. The concatenated features are then integrated, including dimensional alignment of the concatenation position and uniform scaling of the feature values, so that the static feature part and the dynamic feature part have comparable expression scales in the vector space, forming a comprehensive user profile vector.

[0019] It should be noted that the dominant preference label is the content attribute tag with the highest weight extracted, rather than just a single tag. The purpose is to retain all content attribute tags with the same highest weight when multiple weights are tied. The hierarchical labeling is based on the matching relationship between the dominant preference label and the long-term preference setting data in the static attribute data. The long-term preference setting data is stable preference information that users actively set or have accumulated over a long period of time. When the dominant preference label and the long-term preference setting data are consistent, the user is directly assigned to the corresponding hierarchical label. When they are inconsistent, the dominant preference label is the primary factor and the long-term preference setting is the secondary factor for comprehensive hierarchicaling. The time window slides over time, incorporating data from new windows while removing data from the oldest windows. This ensures that the profile is always calculated based on the behavioral data of the most recent consecutive windows. After the window slides, the user's dynamic preference features are recalculated, and the comprehensive profile vector and hierarchical labels are updated accordingly, so that the hierarchical labels can be dynamically adjusted with changes in user behavior.

[0020] Step 2: Profile Tag Matching and Push Execution. Based on user profiles and hierarchical tags, establish mapping rules between hierarchical tags and push content. Match corresponding push content to users according to the mapping rules. Set push rules based on user behavior characteristics. Filter users according to the push rules and execute pushes. Record push response data. In the above embodiments, the specific method for establishing mapping rules between hierarchical tags and push content based on user profiles and hierarchical tags, and matching corresponding push content to users according to the mapping rules, is as follows: For each hierarchical label, obtain the dynamic user preference features of all users under that label, accumulate the weights of each content attribute label, and take the content attribute label corresponding to the highest accumulated weight as the preference content label corresponding to that hierarchical label. Establish a mapping rule between hierarchical labels and push content, and associate the preference content label with the attribute labels marked on the push content in the push content database. For a single user, obtain their current hierarchical label, retrieve the candidate push content associated with that hierarchical label from the push content database according to the mapping rule, calculate the cosine similarity between the user's dynamic preference features and the attribute label vectors of each candidate push content, and select the push content with the highest cosine similarity as the matching result.

[0021] It should be noted that when establishing mapping rules, the dynamic preference features of all users under a hierarchical label are aggregated. All content attribute tags corresponding to the highest accumulated weight are used as the preference content tags for that hierarchical label. If there is a tie for the highest weight, all are retained to fully reflect group preferences. For example, if the accumulated weights for users under a certain hierarchical label show that "sports shoes" and "running shoes" have the highest weights, then both are used as preference tags. After the rules are established, the attribute tags of the pushed content must use the same tag system as the user side. When matching a single user, candidate pushed content is first retrieved from the content library according to the mapping rules corresponding to their hierarchical label. Then, the cosine similarity between the user's dynamic preference features and the attribute tag vectors of each candidate content is calculated, and the highest similarity is selected. For example, if user A has a high weight for "sports shoes" in their dynamic preference features, and there are two types of candidate content (sports shoes and running shoes), the cosine similarity for "sports shoes" is higher, so that content is matched.

[0022] In the above embodiments, the specific method for setting push rules based on user behavior characteristics, filtering users and executing push notifications according to the push rules, and recording push response data is as follows: The system calculates the total frequency of user behavior in each natural hour within each time window, and determines the continuous time period with the highest total frequency as the user's daily cyclic push time window. It also calculates the number of historical clicks by users on each push channel, and determines the push channel with the highest number of clicks as the user's preferred push channel. A daily push frequency limit is set, and the number of pushes received by the user that day is calculated. For users who have completed content matching, if the push content carries a time-sensitive identifier and is currently within the content's validity period, the push time window limit is ignored, and the push is executed when the preferred push channel is available and the number of pushes received that day has not reached the frequency limit. If the push content does not carry a time-sensitive identifier, the push is executed when the current time is within the user's daily cyclic push time window, the preferred push channel is available, and the number of pushes received that day has not reached the frequency limit. The user's click behavior, conversion behavior, closing behavior, and unsubscription behavior after the push are recorded as push response data. Set a waiting period threshold. If the conditions for pushing are not met, the condition will be re-evaluated after the conditions are met. If the push content carries a time-sensitive identifier and the conditions are not met even after the content expires, the push will be abandoned. If the push content does not carry a time-sensitive identifier and the push conditions are not met even after the waiting period threshold is exceeded, the push will be abandoned.

[0023] It should be noted that the daily cyclical push notification window is determined based on the total frequency of user behavior within each natural hour. The purpose is to map historical behavioral patterns into repeatable daily push notification opportunities. By statistically analyzing the frequency of user behavior across different natural hour periods, the system identifies consecutive time slots corresponding to peak activity, ensuring push notifications match user habits. Historical click counts for each push channel determine preferred channels, and user responses to channels are judged based on actual responses rather than the number of push notifications. Time-sensitive content is exempt from push notification window restrictions to ensure timely delivery of limited-time events within their validity period. Content without time-sensitive indicators is limited to push notifications within the specified window to prevent non-urgent pushes from disturbing users at inappropriate times.

[0024] Step 3: Evaluation of push effect and optimization of rules. Based on the recorded push response data, calculate the conversion increment and negative feedback rate. Adjust the time window length and hierarchical tag division according to the changes in conversion increment and negative feedback rate, and optimize the mapping rules and push rules.

[0025] In the above embodiments, the specific method for calculating the conversion increment and negative feedback rate based on the recorded push response data is as follows: After this push notification was executed, the total number of users who received the push, the number of users who converted, the total number of times the push content was closed, and the total number of times users unsubscribed were counted. The conversion rate of this push was obtained by dividing the number of users who converted by the total number of users who received the push. From the users who did not receive the push during this push period, who were in the same tier as the users who received the push and met the push trigger conditions, a group of users equal in number to the number of users who received the push were selected as the control group according to the cosine similarity of the comprehensive profile vector from high to low. The conversion rate of the control group was calculated as the baseline conversion rate. The conversion rate of this push was subtracted from the baseline conversion rate, and the difference was multiplied by the total number of users who received the push to obtain the conversion increment. The number of times the push content was closed and the number of times users unsubscribed were assigned different weights and then weighted and summed to obtain the total number of negative feedbacks. The total number of negative feedbacks was divided by the total number of users who received the push to obtain the negative feedback rate.

[0026] It should be noted that the conversion increment calculation adopts a control group comparison method. The control group is selected from users with the same hierarchical tags as the users who have been pushed to, who meet the push trigger conditions but have not received this push. The cosine similarity of the comprehensive profile vector is used to ensure that the characteristics of the two groups of users are comparable. The conversion rate of the control group reflects the natural conversion level of this group to the same type of content in the same period. The conversion increment calculated in this way can isolate the influence of user group differences and natural conversion fluctuations. The negative feedback rate is calculated by assigning different weights to the closing behavior and the unsubscribing behavior and then summing them. The unsubscribing behavior reflects the user's aversion level significantly higher than the single closing behavior, so it is given a higher calculation weight. Dividing the total number of weighted negative feedbacks by the total number of pushed users can reflect the difference in the severity of negative feedback.

[0027] In the above embodiments, the specific method for adjusting the time window length and hierarchical label division based on changes in conversion increment and negative feedback rate, and optimizing the mapping rules and push rules, is as follows: Compare the conversion increment obtained in this calculation with zero and the negative feedback rate with zero. If the conversion increment is less than zero and the negative feedback rate is equal to zero, adjust the time window length and reclassify the tiered tags. If the conversion increment is greater than zero and the negative feedback rate is equal to zero, keep the current time window length, tiered tag division, mapping rules and push rules unchanged. If the conversion increment is less than zero and the negative feedback rate is greater than zero, the time window length will be adjusted. When dividing the tiered tags, the content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be excluded from the dominant preference tags and included in the individual content blacklist of the corresponding user. They will be removed when the user matches subsequent content. The mapping rules of the corresponding tiered tags will be updated. The content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be removed from the preferred content tags of the tiered tag. At the same time, the daily push frequency limit will be adjusted. If the conversion increment is greater than zero and the negative feedback rate is greater than zero, the time window length remains unchanged. When dividing the tiered tags, the content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be excluded from the dominant preference tags and included in the individual content blacklist of the corresponding user. They will be removed when the user matches content in the future. The mapping rules of the corresponding tiered tags will be updated, and the content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be removed from the preferred content tags of that tiered tag. At the same time, the daily push frequency limit will be adjusted.

[0028] It's important to note that comparing conversion increment and negative feedback rate to zero creates four combinations, each corresponding to different push notification performance states. A conversion increment less than zero indicates the push notification did not bring positive improvement, while a negative feedback rate greater than zero indicates the push notification triggered user aversion. Based on these two signal combinations, a differentiated adjustment strategy is adopted for the time window length. When there is no negative feedback, the window can be expanded to capture longer-term preferences; when negative feedback occurs, the window is expanded while simultaneously optimizing stratification and push rules. When there is positive conversion and negative feedback, the window length is kept constant, adjusting only the content and frequency to avoid unnecessary disturbance to the positive effect. Optimization of individual blacklists and group mapping rules is handled in a stratified manner. Content types that trigger negative feedback are first added to the individual blacklist to ensure that the user will not receive similar content again; simultaneously, they are removed from the preferred content tags in the group mapping rules to reduce the push priority of this type of content at the group level, preventing similar negative feedback from spreading within the stratification. For example, if a user under a certain tier tag significantly reduces or cancels appliance repair-related push notifications, and the system calculates a conversion increase of less than zero and a negative feedback rate of greater than zero within the evaluation period, optimization will involve adding the appliance repair tag to the individual blacklist for each user who triggers negative feedback. Simultaneously, when reclassifying tier tags, "appliance repair" will be excluded from the dominant preference tags and removed from the preference content tags of that tier's mapping rule. Subsequent push notifications for that tier will no longer retrieve this type of candidate content, and the daily push frequency limit will be appropriately reduced.

[0029] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for optimizing precision push of digital marketing based on user behavior portrait, characterized in that, include: Step 1: User Behavior Profile Construction and Update. Collect static attribute data and dynamic behavior data from the entire digital marketing process. Based on the dynamic behavior data, divide the time window into multiple consecutive time windows according to the time of behavior occurrence. Aggregate the dynamic behavior data in each window to obtain the original behavior feature representation. Apply time decay weight to each original behavior feature representation according to the time distance between each time window and the current time to form the user dynamic preference feature. Construct a comprehensive user profile based on the static attribute data and the user dynamic preference feature, and classify users into corresponding hierarchical labels. As time goes by, slide the time window to dynamically update the comprehensive user profile and hierarchical labels. Step 2: Profile Tag Matching and Push Execution. Based on user profiles and hierarchical tags, establish mapping rules between hierarchical tags and push content. Match corresponding push content to users according to the mapping rules. Set push rules based on user behavior characteristics. Filter users according to the push rules and execute pushes. Record push response data. Step 3: Evaluation of push effect and optimization of rules. Based on the recorded push response data, calculate the conversion increment and negative feedback rate. Adjust the time window length and hierarchical tag division according to the changes in conversion increment and negative feedback rate, and optimize the mapping rules and push rules. 2.The method of claim 1, wherein, The static attribute data includes identity attribute data and long-term preference setting data; The dynamic behavioral data includes page view and click data, search query data, dwell and browsing depth data, add-to-cart and order data, push notifications, and close and unsubscribe data. The dynamic behavior data includes the time when the behavior occurred and the content attribute tags of the behavior object. The content attribute tags include content type tags and timeliness indicators. 3.The method of claim 2, wherein, The method involves collecting static attribute data and dynamic behavior data from the entire digital marketing process of users. Based on the dynamic behavior data, multiple consecutive time windows are divided according to the time of behavior occurrence. The dynamic behavior data within each window is aggregated to obtain the original behavior feature representation. According to the time distance between each time window and the current time, a time decay weight is applied to each original behavior feature representation to form the user's dynamic preference features. The specific method is as follows: The system performs missing value imputation and duplicate value removal on dynamic behavior data, normalizes continuous numerical indicators, and performs one-hot encoding on discrete content attribute tags to form standardized dynamic behavior features. A fixed-length time window is set, and each standardized dynamic behavior feature is assigned to a corresponding time window based on the occurrence time of the behavior. Within each time window, a positive initial weight is set according to the behavior type of each behavior feature, while the initial weights for closing and unsubscribing behaviors are set to negative values. The absolute value of the negative weight for unsubscribing behavior is greater than that for closing behavior. The negative weights are directly accumulated into the corresponding associated content attribute tags and are not decayed with the time window. In each time window, the content attribute tags and the corresponding positive initial weight results are integrated to form a window original behavior feature vector consistent with the dimension of the content attribute tags. According to the time distance between each time window and the current time, the time decay weight corresponding to each window is calculated through an exponential decay function, and the decay weight value range is The closer the window to the current time, the greater the decay weight. The decay weight of the current window is 1. The original behavior feature vector of each window is multiplied by the corresponding decay weight, and the feature vectors of all windows after the decay processing are summed up. The cumulative negative weight vector in the whole cycle is superimposed to obtain the user dynamic preference feature.

4. The digital marketing precision push optimization method based on user behavior portrait according to claim 3, characterized in that, The method for constructing a comprehensive user profile based on static attribute data and dynamic user preference features, and dividing users into corresponding hierarchical labels, dynamically updating the comprehensive user profile and hierarchical labels over time by sliding a time window, is as follows: Static attribute data is encoded to obtain static feature vectors. These static feature vectors are then concatenated with user dynamic preference features. The concatenated features are integrated to form a comprehensive user profile vector. All content attribute labels with the maximum weight in the user dynamic preference features are extracted as the user's dominant preference labels. Based on the dominant preference labels and long-term preference settings in the static attribute data, users are assigned to corresponding hierarchical labels. As time progresses, a sliding time window is used to incorporate dynamic behavior data from the new window and remove dynamic behavior data from the oldest window. The user dynamic preference features are then recalculated, and the comprehensive user profile vector and hierarchical labels are updated. 5.The method of claim 1, wherein, The method for establishing mapping rules between user profiles and hierarchical tags and push content, and matching corresponding push content to users according to the mapping rules, is as follows: For each hierarchical label, obtain the dynamic user preference features of all users under that label, accumulate the weights of each content attribute label, and take the content attribute label corresponding to the highest accumulated weight as the preference content label corresponding to that hierarchical label. Establish a mapping rule between hierarchical labels and push content, and associate the preference content label with the attribute labels marked on the push content in the push content database. For a single user, obtain their current hierarchical label, retrieve the candidate push content associated with that hierarchical label from the push content database according to the mapping rule, calculate the cosine similarity between the user's dynamic preference features and the attribute label vectors of each candidate push content, and select the push content with the highest cosine similarity as the matching result.

6. The digital marketing precision push optimization method based on user behavior portrait according to claim 5, characterized in that, The specific method for setting push rules based on user behavior characteristics, filtering users and executing push notifications according to the push rules, and recording push response data is as follows: The system calculates the total frequency of user behavior in each natural hour within each time window, and determines the continuous time period with the highest total frequency as the user's daily cyclic push time window. It also calculates the number of historical clicks by users on each push channel, and determines the push channel with the highest number of clicks as the user's preferred push channel. A daily push frequency limit is set, and the number of pushes received by the user that day is calculated. For users who have completed content matching, if the push content carries a time-sensitive identifier and is currently within the content's validity period, the push time window limit is ignored, and the push is executed when the preferred push channel is available and the number of pushes received that day has not reached the frequency limit. If the push content does not carry a time-sensitive identifier, the push is executed when the current time is within the user's daily cyclic push time window, the preferred push channel is available, and the number of pushes received that day has not reached the frequency limit. The user's click behavior, conversion behavior, closing behavior, and unsubscription behavior after the push are recorded as push response data. Set a waiting period threshold. If the conditions for pushing are not met, the condition will be re-evaluated after the conditions are met. If the push content carries a time-sensitive identifier and the conditions are not met even after the content expires, the push will be abandoned. If the push content does not carry a time-sensitive identifier and the push conditions are not met even after the waiting period threshold is exceeded, the push will be abandoned.

7. The method for optimizing precise digital marketing push based on user behavior profiles according to claim 1, characterized in that, The specific method for calculating the conversion increment and negative feedback rate based on the recorded push response data is as follows: After this push notification was executed, the total number of users who received the push, the number of users who converted, the total number of times the push content was closed, and the total number of times users unsubscribed were counted. The conversion rate of this push was obtained by dividing the number of users who converted by the total number of users who received the push. From the users who did not receive the push during this push period, who were in the same tier as the users who received the push and met the push trigger conditions, a group of users equal in number to the number of users who received the push were selected as the control group according to the cosine similarity of the comprehensive profile vector from high to low. The conversion rate of the control group was calculated as the baseline conversion rate. The conversion rate of this push was subtracted from the baseline conversion rate, and the difference was multiplied by the total number of users who received the push to obtain the conversion increment. The number of times the push content was closed and the number of times users unsubscribed were assigned different weights and then weighted and summed to obtain the total number of negative feedbacks. The total number of negative feedbacks was divided by the total number of users who received the push to obtain the negative feedback rate.

8. The method for optimizing precise digital marketing push based on user behavior profiles according to claim 7, characterized in that, The method for adjusting the time window length and hierarchical tag division based on changes in conversion increment and negative feedback rate, and optimizing mapping and push rules, is as follows: Compare the conversion increment obtained in this calculation with zero and the negative feedback rate with zero. If the conversion increment is less than zero and the negative feedback rate is equal to zero, adjust the time window length and reclassify the tiered tags. If the conversion increment is greater than zero and the negative feedback rate is equal to zero, keep the current time window length, tiered tag division, mapping rules and push rules unchanged. If the conversion increment is less than zero and the negative feedback rate is greater than zero, the time window length will be adjusted. When dividing the tiered tags, the content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be excluded from the dominant preference tags and included in the individual content blacklist of the corresponding user. They will be removed when the user matches subsequent content. The mapping rules of the corresponding tiered tags will be updated. The content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be removed from the preferred content tags of the tiered tag. At the same time, the daily push frequency limit will be adjusted. If the conversion increment is greater than zero and the negative feedback rate is greater than zero, the time window length remains unchanged. When dividing the tiered tags, the content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be excluded from the dominant preference tags and included in the individual content blacklist of the corresponding user. They will be removed when the user matches content in the future. The mapping rules of the corresponding tiered tags will be updated, and the content attribute tags corresponding to the push content types that trigger unsubscribe or close behavior will be removed from the preferred content tags of that tiered tag. At the same time, the daily push frequency limit will be adjusted.

Images