Digital marketing behavior forgetting and user restart recommendation method and system

Abstract

This invention discloses a method and system for recommending digital marketing behavior forgetting and user restart, relating to the field of digital marketing management. By analyzing user behavior on a platform, this invention quantifies users' long-term interests in various topics and assesses interest decay, identifies topics with high historical preference but declining current interest, and generates targeted recall content for personalized push notifications, thereby achieving interest activation and behavior restart. Through this invention, user re-engagement and content interaction rates can be improved, increasing the conversion efficiency of marketing activities. Simultaneously, it optimizes the push notification rhythm and content priority, reduces user fatigue, and maximizes the long-term value of platform users and enhances the overall precision marketing.

Description

Technical Field

[0001] This invention relates to the field of digital marketing management, specifically a method and system for recommending digital marketing behavior forgetting and user restart. Background Technology

[0002] In the field of digital marketing, recommendation systems and precision marketing have become important technical means to improve user engagement and conversion rates. Existing technologies typically rely on user behavior data, interest models, and collaborative filtering algorithms. By analyzing users' browsing, clicking, saving, and purchasing records, they model user preferences and distribute content through in-app recommendation modules, push notifications, or personalized advertising slots. User interest models generally include two categories: long-term interests, reflecting the accumulated preferences of users' historical behavior, and short-term interests, reflecting the recent trends in user behavior. Collaborative filtering is divided into user-based collaborative filtering and item-based collaborative filtering. User-based collaborative filtering makes recommendations by finding user groups with similar interests, while item-based collaborative filtering generates recommendation lists based on the similarity between content from users' historical interactions. In addition, content filtering and contextual recommendation technologies are also widely used to improve the relevance and personalization of recommendations.

[0003] However, existing technologies still have some limitations in practical applications. First, most recommender systems assume that user interests are static or change slowly, failing to fully capture the natural decay of interests over time. Therefore, when a user's interest in certain content diminishes, recommended content lacks targeting, easily leading to a decrease in user activity. Second, existing methods typically rely on simple weighting or sliding window strategies to integrate historical and short-term interests, lacking fine-grained modeling of the interest decay process, making it difficult to accurately determine when and how to reactivate a user's potential interests. Furthermore, the ranking and frequency management of recommended content in existing technologies often only consider immediate click-through rates or conversion rates, ignoring the psychological patterns of user memory and attention changes over time, resulting in some high-potential content failing to effectively reach users. These shortcomings indicate room for improvement in dynamic interest modeling, decay identification, and personalized recall pushes, requiring new methods to more accurately respond to changes in user behavior and improve marketing effectiveness. Summary of the Invention

[0004] This invention proposes a method and system for recommending digital marketing behavior forgetting and user restart, aiming to solve the problem of difficulty in accurately capturing dynamic changes in user interests in existing digital marketing technologies. By improving interest modeling and behavior analysis methods, the system can more accurately identify the state of user interest decay, optimize the generation and sorting of recall content, and achieve targeted and efficient content push, thereby improving user activity, repurchase rate and overall marketing conversion effect.

[0005] One method for recommending digital marketing behavior forgetting and user restart includes the following steps:

[0006] S1. Obtain user behavior logs on the platform and record the topic tags and timestamps of each behavior. Combine the topic tags and corresponding users to construct user-topic behavior sequences. Based on the user-topic behavior sequences, quantify and accumulate each user's long-term interest in each topic to obtain a baseline interest score that represents the intensity of the user's historical interest in the topic.

[0007] Specifically, every user action on the platform carries time information and topic association. These action logs are first processed in a refined manner. The system organizes data according to the user-topic dimension, assigning different weights to different types of behavior (browsing, clicking, dwelling, saving, sharing, purchasing, unfollowing). A time decay function is introduced to handle the timeliness of the behavior, forming a weighted cumulative algorithm. For example, purchasing behavior has a higher weight than browsing, and recent behavior has a higher weight than older behavior. After weighted accumulation, each user's long-term interest value on each topic can be obtained; this value is then normalized to form a baseline interest score matrix.

[0008] S2. Calculate the interest retention rate of each user for each topic based on the user-topic behavior sequence and timestamp, and calculate the user's forgetting rate based on the calculated interest retention rate;

[0009] Specifically, based on the forgetting curve theory, an exponential-power-law hybrid model is used to calculate the retention of interest. The model takes the time difference between the user's most recent topic behavior and the current time as input. The exponential part reflects the recent rapid forgetting effect, and the power-law part reflects the long-term slow forgetting effect. It can simultaneously capture the short-term and long-term characteristics of user interest decay. The calculation of interest retention reflects the time factor and is dynamically adjusted by combining behavior type weights, behavior frequency, and historical baseline interests. This allows the model to quantify the user's memory retention status for each topic and further calculate the degree of forgetting to measure the magnitude of interest decay.

[0010] S3. Calculate the current interest score for each user on each topic based on the baseline interest score and interest retention rate; and combine the interest score with a preset threshold to determine the user's interest decay status, filter and identify topics with declining interest but high historical interest, and form a recall candidate pool.

[0011] Specifically, based on interest retention and baseline interest, the current interest score is calculated through weighted fusion, and the current interest and historical interest are judged by a preset threshold. Topics with high historical interest but significantly declining current interest are identified, and these user-topic pairs are included in the recall candidate pool. The recall candidate pool stores basic interest parameters, recent behavior time, interest retention, and historical interest weight fields.

[0012] S4. Generate candidate content for recall based on user-topic pairs in the recall candidate pool. Prioritize the candidate content for recall based on interest score, forgetting level and baseline interest score. Generate a push sequence based on user fatigue and push frequency limit.

[0013] Specifically, after the candidate pool is generated, a set of candidate content for each user-topic pair is constructed. Priority is calculated through weighted averages, combining the current interest score, interest retention rate, and baseline interest score to form a comprehensive ranking index. User fatigue, push frequency, and content timeliness constraints can be incorporated into the priority calculation process. The generated push sequence reflects the interest activation potential and avoids user push overload. Candidate content is sorted by priority to form a personalized, directly executable push sequence of recall content.

[0014] S5. Push recall content to users according to the push sequence.

[0015] Furthermore, it also includes a digital marketing behavior forgetting and user restart recommendation system, which is implemented based on any of the aforementioned digital marketing behavior forgetting and user restart recommendation methods, including:

[0016] The data acquisition module obtains user behavior logs on the platform and records the topic tags and timestamps for each behavior.

[0017] The behavior sequence construction module is used to construct user-topic behavior sequences based on topic tags and corresponding users, and to construct user-topic behavior sequences based on user-topic behavior sequences;

[0018] The baseline interest score calculation module is used to quantify and accumulate each user's long-term interest in each topic to obtain a baseline interest score that represents the intensity of the user's historical interest in the topic.

[0019] The forgetting level calculation module is used to calculate the interest retention of each user on each topic based on the user-topic behavior sequence and timestamp, and to calculate the user's forgetting level based on the calculated interest retention.

[0020] The recall candidate pool construction module is used to calculate the current interest score of each user for each topic based on the baseline interest score and interest retention rate; and to determine the user's interest decay status by combining the interest score with a preset threshold, and to filter and identify topics with declining interest but high historical interest to form the recall candidate pool.

[0021] The push sequence generation module generates recall candidate content based on user-topic pairs in the recall candidate pool. It prioritizes the recall candidate content by combining interest scores, forgetting level, and baseline interest scores, and generates a push sequence by combining user fatigue and push frequency limits.

[0022] The memory content push module pushes memory content to users according to the push sequence.

[0023] Specifically, the system acquires user behavior logs, including behavior type, topic tags, and timestamps, through a data acquisition module. These logs are then organized into user-topic behavior sequences by a behavior sequence construction module, providing structured data for interest analysis. A baseline interest score calculation module assigns weights to different behavior types and introduces time decay to accumulate and generate each user's long-term interest intensity across topics. A forgetting rate calculation module analyzes interest retention based on an exponential-power-law hybrid model, mapping it to the degree of user interest decay. A recall candidate pool construction module integrates current interest scores and historical interest thresholds to filter user-topic pairs with declining interest but high historical preference, forming a candidate pool. A push sequence generation module calculates content priority based on interest scores, forgetting rate, and user fatigue, and generates personalized push sequences. A recall content push module delivers candidate content to users through recommendation slots, message notifications, or advertising channels according to the sequence, thereby activating user interest and restarting behavior during the interest decay phase. The entire system forms a complete data-driven closed loop.

[0024] The beneficial effects of the invention are:

[0025] This invention effectively identifies the natural decay trend of user interests over time and accurately generates recall content based on historical preferences, thereby increasing user re-engagement with platform content. Through comprehensive analysis of user interest retention and forgetting levels, the system can scientifically regulate the priority and frequency of pushed content, avoiding user fatigue while ensuring the effective activation of high-potential interests. This overall mechanism not only improves the accuracy and personalization of recommendations but also significantly enhances user activity and the overall marketing conversion effect of the platform, maximizing the utilization of user behavior value and optimizing long-term customer relationship maintenance. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a method for recommending digital marketing behavior forgetting and user restart according to an embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, and not all of them. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0029] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0030] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or machine that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or machine. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or machine that includes said element.

[0031] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0032] Example 1

[0033] Among them, such as Figure 1 A method for recommending digital marketing behavior forgetting and user restart, including the following steps:

[0034] S1. Obtain user behavior logs on the platform and record the topic tags and timestamps of each behavior. Combine the topic tags and corresponding users to construct user-topic behavior sequences. Based on the user-topic behavior sequences, quantify and accumulate each user's long-term interest in each topic to obtain a baseline interest score that represents the intensity of the user's historical interest in the topic.

[0035] S2. Calculate the interest retention rate of each user for each topic based on the user-topic behavior sequence and timestamp, and calculate the user's forgetting rate based on the calculated interest retention rate;

[0036] S3. Calculate the current interest score for each user on each topic based on the baseline interest score and interest retention rate; and combine the interest score with a preset threshold to determine the user's interest decay status, filter and identify topics with declining interest but high historical interest, and form a recall candidate pool.

[0037] S4. Generate candidate content for recall based on user-topic pairs in the recall candidate pool. Prioritize the candidate content for recall based on interest score, forgetting level and baseline interest score. Generate a push sequence based on user fatigue and push frequency limit.

[0038] S5. Push recall content to users according to the push sequence.

[0039] Furthermore, step S1 specifically includes the following sub-steps:

[0040] S101. Collect user behavior logs on the platform according to the platform data interface. The behavior logs include the recorded user behavior type, corresponding content theme and the timestamp of the behavior. Construct a user-theme behavior sequence based on the user behavior type, corresponding content theme and the timestamp of the behavior.

[0041] S102. Based on the user-topic behavior sequence, classify the behaviors according to the topic to obtain a subset of user behaviors for each topic;

[0042] S103. Based on the behavioral subset, the behaviors of each topic are weighted and accumulated to calculate the preliminary interest score;

[0043] S104. Normalize the initial interest scores, and establish a baseline interest score matrix for each topic based on the normalized interest scores.

[0044] Specifically, in a digital marketing environment, to achieve quantitative modeling of user interests, it is necessary to comprehensively collect user behavior logs through platform data interfaces. The collected behavioral data includes not only the user's behavior type (clicks, browsing, favorites, sharing), but also the content theme corresponding to each behavior and the timestamp of the behavior. Based on the raw data, a user-topic behavior sequence is constructed, structurally mapping the user's historical behavior on different topics. This transforms fragmented, asynchronous user operations into quantifiable behavioral trajectories, enabling the characterization and tracking of each user's interest evolution across different topics. After constructing the behavior sequence, user behavior records are categorized by topic, obtaining behavioral subsets under each topic. This distinguishes the role of different types of behavior in interest formation; for example, browsing may contribute limitedly to interest, while favorites or sharing usually represent stronger interest signals. Through topic classification, interest calculation combines multiple dimensions of behavioral characteristics to reflect the depth of user preferences. Furthermore, a weighted accumulation of behavioral subsets for each topic yields a preliminary interest score. The principle of weighted accumulation is to assign weights to different behavior types and time dimensions, ensuring that recent, high-frequency, and highly interactive behaviors contribute more, while earlier behaviors gradually diminish in contribution. The interest score measures a user's overall activity level on a particular topic, reflecting the dynamics and hierarchy of interest formation. Further, the preliminary interest scores are normalized to eliminate biases caused by differences in user behavior magnitude or topic activity levels, and a user-topic baseline interest score matrix is ​​constructed. This matrix presents the long-term interest distribution of users across various topics, reflecting the underlying preference structure. Through the user-topic baseline interest score matrix, topics that users once highly interacted with but may have gradually forgotten are quantitatively identified, enabling the platform to reactivate these interests in subsequent recommendations, thereby reactivating user interests and optimizing personalized recommendations.

[0045] For example, a sample data structure for behavior logs is provided for reference:

[0046] user_id (string / int);

[0047] content_id (string / int);

[0048] topic_id (string / int);

[0049] event_type (enum: view / click / like / share / fav / purchase / unfollow / … );

[0050] event_weight (float);

[0051] ts (timestamp);

[0052] Furthermore, the behavior of each topic is weighted and accumulated using either the EWMA method or the original weighted accumulation method. An exemplary flow of the original weighted accumulation method is given below:

[0053] Calculate the raw score for a subset of user u's behaviors in topic j:

[0054] ;

[0055] ;

[0056] Among them, the This represents the original baseline interest score (unnormalized) of user u on topic j. Let i represent the set of behaviors of user u on topic j, where i represents the behavior index. This represents the weight of the i-th behavior, used to measure the contribution of behavior type to interest. The timestamp representing the occurrence of the i-th action, the Indicates the current timestamp, the Indicates a time interval, the The time decay function is used to adjust the effective contribution of the behavior. This represents the exponential decay rate parameter.

[0057] Furthermore, the user behavior types recorded in the behavior log include at least browsing, clicking, staying, collecting, sharing, purchasing, and unfollowing.

[0058] Furthermore, step S2 specifically includes the following sub-steps:

[0059] S201. Based on the forgetting curve theory, construct an exponential-power law hybrid model to quantify the natural decay trend of users' interest in various topics over time;

[0060] S202. For each user in each topic, retrieve the user-topic behavior sequence to obtain the time of the user's most recent behavior in that topic;

[0061] S203. Compare the current time with the time of the most recent action to obtain the time interval between the two; and use the time interval as input to the exponential-power-law hybrid model to calculate the interest retention degree;

[0062] S204. Calculate the user's forgetting level based on interest retention rate.

[0063] Specifically, in the modeling of interest retention and forgetting, a computational framework is established to characterize the changing patterns of user interests over time. Based on the forgetting curve theory, interest does not remain constant over time, but rather exhibits a trend of rapid decay and slow forgetting. To accurately reflect this phenomenon, a hybrid model combining exponential and power-law approaches is adopted to sensitively capture the rapid decline in interest in the short term, while preserving the user's potential stable preferences in the long term. In practical implementation, it is necessary to retrieve the time of the most recent interaction for each topic from the user's behavior records. This time point is regarded as the starting position for timing interest intensity, used to measure the user's latest engagement on that topic. If there is no behavior data for a certain topic for a long period of time, a default low value or exclusion can be set to ensure the stability and accuracy of the model. By comparing the current time with this time point, the time interval between the two can be obtained. The time interval is used as a key input variable to measure interest retention, intuitively reflecting the user's silence duration on that topic.

[0064] Furthermore, the time interval is input into the hybrid decay model to obtain the interest retention rate. This value falls within a relative range and represents the degree of residual interest a user has on a particular topic. The level of interest retention rate depends not only on the distance between the user's most recent interaction and the current situation but also on the weighting of short-term and long-term factors, thus ensuring the model's adaptability to different types of topics and users with varying activity levels. For topics with recent high levels of interaction, the interest retention rate is assessed as high, indicating that the user may still be actively engaged. Conversely, for topics with long-term inactivity, the interest retention rate gradually decreases, indicating that the user's attention to the topic is gradually fading. Further, the user's forgetting rate is derived. Forgetting rate, as the dual concept of interest retention rate, is used to characterize the level of waning interest in a particular topic. The higher the retention rate, the lower the forgetting rate; conversely, the lower the retention rate, the higher the forgetting rate. This complementary relationship clearly distinguishes between interest maintenance and interest loss numerically. By jointly modeling retention rate and forgetting rate, dynamic monitoring of the user's interest status can be achieved.

[0065] Furthermore, the forgetting curve theory was first proposed by the German psychologist Hermann Ebbinghaus in the 19th century to reveal the natural decay of human memory over time. This theory, through experiments, found that forgetting after learning new knowledge does not occur linearly and uniformly, but rather exhibits a two-stage characteristic of "rapid-slow": in the initial period after memory formation, the rate of forgetting is extremely fast, with most memories lost in a short time; as time goes on, the rate of forgetting gradually slows down, and the remaining memories are retained at a lower rate for a longer period. This pattern indicates that the human memory system is highly sensitive in the short term, but still has a stable residual effect in the long term. Based on this theory, the forgetting curve is usually formally described as a function model that decreases over time. Exponential functions can well fit the rapid decay process in the short term, while power-law functions better fit the tail effect of the slow decay of long-term memory. A single model often cannot simultaneously capture both short-term and long-term characteristics; therefore, in practical modeling, a hybrid modeling approach combining exponential and power-law methods is often used to obtain a more accurate characterization of memory decay based on psychological and behavioral empirical evidence.

[0066] Based on the aforementioned principles, this embodiment analogizes user behavior and interests on the platform to "memory traces." User interactions on a particular topic (such as browsing, clicking, saving, and purchasing) can be seen as a process of "memory reinforcement." When a user does not engage in related behavior for a period of time, their interest gradually declines, similar to the natural forgetting of memories. By introducing the mathematical concept of the forgetting curve, the solution can dynamically adjust the retention rate of interest based on the time interval between user behaviors, thereby avoiding the problem of static cumulative models overestimating or underestimating user interest.

[0067] Furthermore, in step S201, the exponential-power hybrid model is specifically represented as follows:

[0068] ;

[0069] Among them, the Indicates user The interest retention rate for topic j, i.e., the proportion of time t during which a user retains interest in that topic, ranges from [0,1]. This represents the exponential decay weighting coefficient, used to control the proportion of short-term interest contributing to the overall retention rate. The exponential decay rate parameter is used to control the rate of decay of recent behavior. Indicates the current timestamp, the This represents the timestamp of user u's most recent action in topic j. This represents the power-law decay exponent, used to indicate the slow decay of interest in long-term memory.

[0070] Furthermore, the specific calculation process of step S203 is expressed as follows:

[0071] Calculation time interval:

[0072] ;

[0073] Will Substituting into the formula for the exponential-power-law hybrid model, we can calculate the interest retention degree, i.e.:

[0074] ;

[0075] Among them, the Indicates user The interest retention rate for topic j, i.e., the proportion of time t during which a user retains interest in that topic, ranges from [0,1]. This represents the exponential decay weighting coefficient, used to control the proportion of short-term interest contributing to the overall retention rate. The exponential decay rate parameter is used to control the rate of decay of recent behavior. Indicates the current timestamp, the This represents the timestamp of user u's most recent action in topic j. This represents the power-law decay exponent, used to indicate the slow decay of interest in long-term memory.

[0076] Furthermore, in step S204, the specific calculation process for calculating the user's degree of forgetfulness is as follows:

[0077] ;

[0078] Among them, the This indicates the degree to which user u has forgotten topic j. Indicates user Maintaining interest in topic j.

[0079] Furthermore, step S3 specifically includes the following sub-steps:

[0080] S301. Calculate the user's current interest score on the corresponding topic based on the baseline interest score and interest retention rate;

[0081] S302. By setting the current interest threshold, determine the degree of user interest decay;

[0082] S303. Determine the intensity of a user's historical interests by setting a historical interest threshold;

[0083] S304. Combining the results of the interest decay determination with the results of the historical interest intensity determination, select user-topic pairs that meet the preset conditions to enter the memory candidate pool.

[0084] Specifically, in the process of determining interest decay and constructing the recall candidate pool, the baseline interest score and interest retention obtained from previous steps are combined to form an interest score that dynamically reflects the user's current state. The baseline interest score, as the user's long-term accumulated preference on the topic, reflects the overall contribution of historical behavior; while the interest retention characterizes the decay effect of this preference over time. By combining the two, the user's actual level of interest in the specific topic at the current moment is calculated. After obtaining the current interest score, a reasonable threshold is set to determine whether the score is in the decay range. For example, for frequently interacted topics, the threshold is relatively higher to enhance the sensitivity to decay trends; while for low-frequency topics, a lower threshold is set to avoid prematurely judging interest decay due to short-term inactivity. By comparing with the threshold, it is determined whether the user's current interest in the topic has significantly decreased, thus obtaining the result of interest decay determination.

[0085] Furthermore, the historical interest intensity of users on this topic is independently assessed. This assessment is based on baseline interest scores and combined with a historical interest threshold. By identifying topics that have shown high attention in long-term behavior, even if current interest has declined, they still possess potential "reactivation" value. In other words, the historical interest threshold is set to distinguish between "weakly correlated incidental behavior" and "stable deep preferences," avoiding misjudging irrelevant topics as objects requiring recall activation. By combining the assessment results of current interest decay and historical interest intensity, user-topic pairs that meet preset conditions are selected and included in the recall candidate pool. Topics entering the candidate pool have two typical characteristics: firstly, they have historically had high baseline interest, indicating that users have previously invested significant attention or interaction on these topics; secondly, their current interest level has significantly declined, indicating a decrease in users' willingness to engage proactively. These types of topics, characterized by "strong historical interest but current decline," are suitable for stimulating user re-engagement through recall-based content pushes.

[0086] Furthermore, for the recall candidate pool, by determining the degree of interest decay and historical interest levels of users on each topic, user-topic pairs that meet preset conditions are included in the candidate pool, forming the input set for subsequent content generation and sorting processing. Each user-topic pair in the candidate pool carries interest decay status, historical interest scores, and relevant time information. In addition, by storing and indexing user-topic pairs in the candidate pool, various parameters can be directly referenced for sorting and filtering in subsequent processing stages, without having to repeatedly calculate interest determination or retrieve the original behavior sequence.

[0087] The user-topic sequence is a time-series data structure used to record each user's complete behavioral history on each topic, including behavior type, timestamp, and frequency. It reflects the user's behavioral trajectory and interest change trends on specific topics, and is used to calculate interest retention, forgetting rate, and baseline interest score, forming the basis for dynamic interest modeling. The user-topic pair, on the other hand, is an entity extracted from the user-topic sequence after interest decay analysis and threshold filtering. Only combinations of users and topics with significant current interest decay but high historical interest are retained to form a recall candidate pool, providing structured objects for recall content generation and personalized push notifications. Therefore, the user-topic sequence is a complete representation of the original behavioral data, while the user-topic pair is a set of high-potential targets after interest decay determination and filtering.

[0088] Furthermore, the filtering rules for selecting user-topic pairs that meet the preset conditions to enter the memory candidate pool are specifically expressed as follows:

[0089] ;

[0090] Among them, the Represents the candidate pool of memories, the Representing user-topic pairs, the This represents the current interest score of user u in topic j. Indicates the current interest threshold, the This represents the baseline interest score of user u in topic j. This represents the historical interest threshold.

[0091] Furthermore, step S4 specifically includes the following sub-steps:

[0092] S401. Based on the memory candidate pool, generate a set of memory candidate content for each user's corresponding theme;

[0093] S402. For each candidate content, calculate the priority by weighting the current interest score, interest retention rate, and baseline interest score;

[0094] S403. For each user, sort the candidate memories from high to low according to the adjusted priority, and generate a personalized memory content push sequence.

[0095] Specifically, in the process of generating and pushing memory content, a candidate pool is used as a foundation to map each user's user-topic pair in the pool to a corresponding set of memory candidate content. Each candidate content carries parameter information related to that user-topic pair, including the current interest score, interest retention rate, and baseline interest score. Priority quantification is performed on each candidate content, with the priority calculated based on three parameters: the current interest score reflects the user's immediate interest status on the topic; the interest retention rate characterizes the degree of decay of the user's interest over time; and the baseline interest score reflects the user's long-term cumulative preference on the topic. A unified numerical index is obtained by weighting these three parameters to distinguish the relative importance of different candidate content in memory pushes. The weighting method is adjusted according to business needs and topic characteristics.

[0096] Furthermore, the candidate memories for each user are sorted from high to low according to the calculated priority to form the final personalized memory content push sequence. This sorting process is based solely on the parameter information in the candidate content set and does not involve external recommendation rules or effect optimization logic.

[0097] Furthermore, the push notifications for meeting content include products or topics that users have interacted with highly in the past, as well as related new products, promotional activities, or dynamic advertising information. The system will calculate the weighted priority of candidate content based on the user's current interest score, interest retention rate, and baseline interest score. At the same time, it will optimize the sorting by taking into account user fatigue and push frequency limits to ensure that the content can effectively activate the user's interest when it reaches the user, while avoiding repetition or excessive interference. The push format can be presented through in-app recommendation modules, message notifications, emails, or social media advertisements.

[0098] Example 2

[0099] Furthermore, as a preferred embodiment of the above embodiments, a digital marketing behavior forgetting and user restart recommendation system is proposed. This system is implemented based on the digital marketing behavior forgetting and user restart recommendation method described in any one of the above embodiments, wherein it includes:

[0100] The data acquisition module obtains user behavior logs on the platform and records the topic tags and timestamps for each behavior.

[0101] The behavior sequence construction module is used to construct user-topic behavior sequences based on topic tags and corresponding users, and to construct user-topic behavior sequences based on user-topic behavior sequences;

[0102] The baseline interest score calculation module is used to quantify and accumulate each user's long-term interest in each topic to obtain a baseline interest score that represents the intensity of the user's historical interest in the topic.

[0103] The forgetting level calculation module is used to calculate the interest retention of each user on each topic based on the user-topic behavior sequence and timestamp, and to calculate the user's forgetting level based on the calculated interest retention.

[0104] The recall candidate pool construction module is used to calculate the current interest score of each user for each topic based on the baseline interest score and interest retention rate; and to determine the user's interest decay status by combining the interest score with a preset threshold, and to filter and identify topics with declining interest but high historical interest to form the recall candidate pool.

[0105] The push sequence generation module generates recall candidate content based on user-topic pairs in the recall candidate pool. It prioritizes the recall candidate content by combining interest scores, forgetting level, and baseline interest scores, and generates a push sequence by combining user fatigue and push frequency limits.

[0106] The memory content push module pushes memory content to users according to the push sequence.

[0107] Specifically, the data acquisition module acquires user behavior logs on the platform. Each log entry records the behavior type, corresponding topic tags, and the timestamp of the behavior, providing the raw data foundation for subsequent analysis. The behavior sequence construction module processes the collected data, organizing user-topic behavior sequences according to the correspondence between users and topic tags. It integrates all behaviors of each user on each topic in chronological order, maintaining the integrity of behavior type and time information, providing structured input for interest modeling and dynamic analysis. The baseline interest score calculation module assigns weights to different behavior types based on the user-topic behavior sequences and introduces a time decay function to perform weighted accumulation of behaviors. The product sum and normalization process calculates the long-term interest intensity of each user on each topic, forming a baseline interest score matrix to represent historical interests. This matrix reflects users' cumulative preferences and behavioral differences, providing a reference for subsequent dynamic interest calculations. The forgetting curve calculation module, based on forgetting curve theory and an exponential-power-law hybrid model, calculates the interest retention of each user on each topic. By analyzing the interval between the current time and the most recent behavior time, and combining recent behavioral activity with long-term baseline interests, it maps interest retention to the user's forgetting level, quantifying the decline in user interest in a specific topic and providing a basis for judging interest decay. The recall candidate pool construction module, based on the current... The interest score calculation formula integrates the baseline interest score with interest retention rate to generate the user's current interest score for each topic. It also determines the user's interest decay status through a preset threshold and filters topics with declining interest but high historical preference by combining historical interest intensity. User-topic pairs meeting these criteria are included in the recall candidate pool. This module not only stores interest scores and decay information but also retains parameters such as recent behavior time and interest retention rate, providing structured data support for recall content generation. The push sequence generation module generates a set of recall candidate content for each user-topic pair in the candidate pool. Through weighted calculation, it combines the current interest score, forgetting level, and baseline interest score to form a comprehensive optimization. The system prioritizes candidate content and incorporates user fatigue and push frequency constraints into the calculation, sorting candidate content from high to low priority to form a personalized recall content push sequence, achieving scientific sorting and precise coverage of push content. Based on the generated push sequence, the recall content push module pushes recall content to users through channels such as in-app recommendations, message notifications, emails, or social media ads, thereby effectively activating potential preferences and promoting user re-engagement during the interest decay stage. The entire system forms a complete technical closed loop through data collection, interest modeling, decay analysis, candidate pool construction, priority sorting, and precise push, realizing the systematic management and execution of digital marketing behavior forgetting and user restart.

[0108] Example 3

[0109] Furthermore, as a preferred embodiment of the above-described Example 1, an application of a digital marketing behavior forgetting and user restart recommendation method is proposed. Specifically, this method is applied in digital marketing scenarios:

[0110] In digital marketing scenarios, user behavior logs are collected from the platform, including browsing, clicking, dwell time, saving, sharing, purchasing, and unfollowing behaviors. The corresponding product categories and the time of each behavior are also recorded. Based on the behavioral data from the past 30 days, user behavior in the "smartwatch" product category is weighted and accumulated: 20 views, 3 saves, and 1 purchase, resulting in a baseline interest score of 0.82. For the "sports shoes" product category, 8 views, 1 save, and 0 purchases, the baseline interest score is 0.35.

[0111] Based on the forgetting curve theory, an exponential-power-law hybrid model was used to characterize the natural decay of user interest over time. The most recent behavior in the "smartwatch" category occurred 5 days ago, and the most recent behavior in the "sports shoe" category occurred 18 days ago. After inputting the interval between the current time and the most recent behavior time into the model, the interest retention rate for "smartwatch" was 0.70, and the interest retention rate for "sports shoe" was 0.30, corresponding to forgetting rates of 0.30 and 0.70, respectively.

[0112] The current interest score is calculated by weighting the baseline interest score with the interest retention rate: "Smartwatch" is 0.82 × 0.70 = 0.574, and "Sneakers" is 0.35 × 0.30 = 0.105. Setting the current interest threshold to 0.3 and the historical interest threshold to 0.3, the "Smartwatch" category is determined not to have significantly diminished interest and is therefore not included in the recall candidate pool; the "Sneakers" category has a current interest below the threshold but a historical interest above the threshold, and therefore is included in the candidate pool. The user-product category pairs stored in the candidate pool include category identifier, historical interest score, current interest score, interest retention rate, and recent behavior time, providing structured input for the generation of subsequent recall-based marketing content.

[0113] A set of recall-based marketing content candidates is generated based on the candidate pool. The priority of each candidate is calculated using a weighted average. Taking the "Sneakers" category as an example, with a current interest score of 0.105, an interest retention rate of 0.30, and a baseline interest score of 0.35, the weighted average yields a priority of 0.23. The candidate content is then sorted according to priority to generate a personalized recall-based marketing content push sequence. In this sequence, the "Sneakers" category has the highest priority, and the remaining candidate categories are arranged in order of their calculated priorities for subsequent targeted pushes of discounts, promotions, or new product information.

[0114] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for recommending digital marketing behavior forgetting and user restart, characterized in that, Includes the following steps: S1. Obtain user behavior logs on the platform and record the topic tags and timestamps of each behavior. Combine the topic tags and corresponding users to construct user-topic behavior sequences. Based on the user-topic behavior sequences, quantify and accumulate each user's long-term interest in each topic to obtain a baseline interest score that represents the intensity of the user's historical interest in the topic. S2. Calculate the interest retention rate of each user for each topic based on the user-topic behavior sequence and timestamp, and calculate the user's forgetting rate based on the calculated interest retention rate; S3. Calculate the user's current interest score for the corresponding topic based on the baseline interest score and interest retention rate; determine the degree of user interest decay by setting a current interest threshold; determine the intensity of user's historical interest by setting a historical interest threshold; combine the results of the interest decay determination and the historical interest intensity determination to select user-topic pairs that meet the preset conditions to enter the recall candidate pool. S4. Generate candidate content for recall based on user-topic pairs in the recall candidate pool. Prioritize the candidate content for recall based on interest score, forgetting level and baseline interest score. Generate a push sequence based on user fatigue and push frequency limit. S5. Push recall content to users according to the push sequence; Specifically, step S2 includes the following sub-steps: S201. Based on the forgetting curve theory, construct an exponential-power law hybrid model to quantify the natural decay trend of users' interest in various topics over time; S202. For each user in each topic, retrieve the user-topic behavior sequence to obtain the time of the user's most recent behavior in that topic; S203. Compare the current time with the time of the most recent action to obtain the time interval between the two; and use the time interval as input to the exponential-power-law hybrid model to calculate the interest retention degree; S204. Calculate the user's forgetting level based on interest retention rate; In step S201, the exponential-power law hybrid model is specifically represented as follows: ； Among them, the Indicates user The interest retention rate for topic j, i.e., the proportion of time t during which a user retains interest in that topic, ranges from [0,1]. This represents the exponential decay weighting coefficient, used to control the proportion of short-term interest contributing to the overall retention rate. The exponential decay rate parameter is used to control the rate of decay of recent behavior. Indicates the current timestamp, the This represents the timestamp of user u's most recent action in topic j. This represents the power-law decay exponent, used to indicate the slow decay of interest in long-term memory.

2. The method for recommending digital marketing behavior forgetting and user restart as described in claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S101. Collect user behavior logs on the platform according to the platform data interface. The behavior logs include the recorded user behavior type, corresponding content theme and the timestamp of the behavior. Construct a user-theme behavior sequence based on the user behavior type, corresponding content theme and the timestamp of the behavior. S102. Based on the user-topic behavior sequence, classify the behaviors according to the topic to obtain a subset of user behaviors for each topic; S103. Based on the behavioral subset, the behaviors of each topic are weighted and accumulated to calculate the preliminary interest score; S104. Normalize the initial interest scores, and establish a baseline interest score matrix for each topic based on the normalized interest scores.

3. The method for recommending digital marketing behavior forgetting and user restart as described in claim 2, characterized in that, The user behavior log records at least the following user behavior types: browsing, clicking, staying, saving, sharing, purchasing, and unfollowing.

4. The method for recommending digital marketing behavior forgetting and user restart as described in claim 1, characterized in that, The specific calculation process for step S203 is as follows: Calculation time interval: ； Will Substituting into the formula for the exponential-power-law hybrid model, we can calculate the interest retention degree, i.e.: 。 5. The method for recommending digital marketing behavior forgetting and user restart as described in claim 1, characterized in that, In step S204, the specific calculation process for calculating the user's degree of forgetfulness is as follows: ； Among them, the This indicates the degree to which user u has forgotten topic j.

6. The method for recommending digital marketing behavior forgetting and user restart as described in claim 1, characterized in that, The filtering rules for selecting user-topic pairs that meet the preset conditions to enter the memory candidate pool are specifically expressed as follows: ； Among them, the Represents the candidate pool of memories, the Representing user-topic pairs, the This represents the current interest score of user u in topic j. Indicates the current interest threshold, the This represents the baseline interest score of user u in topic j. This represents the historical interest threshold.

7. The digital marketing behavior forgetting and user restart recommendation method as described in claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S401. Based on the memory candidate pool, generate a set of memory candidate content for each user's corresponding theme; S402. For each candidate content, calculate the priority by weighting the current interest score, interest retention rate, and baseline interest score; S403. For each user, sort the candidate memories from high to low according to the adjusted priority, and generate a personalized memory content push sequence.

8. A digital marketing behavior forgetting and user restart recommendation system, the system being implemented based on the digital marketing behavior forgetting and user restart recommendation method described in any one of claims 1-7, characterized in that, include: The data acquisition module obtains user behavior logs on the platform and records the topic tags and timestamps for each behavior. The behavior sequence construction module is used to construct user-topic behavior sequences based on topic tags and corresponding users, and to construct user-topic behavior sequences based on user-topic behavior sequences; The baseline interest score calculation module is used to quantify and accumulate each user's long-term interest in each topic to obtain a baseline interest score that represents the intensity of the user's historical interest in the topic. The forgetting level calculation module is used to calculate the interest retention of each user on each topic based on the user-topic behavior sequence and timestamp, and to calculate the user's forgetting level based on the calculated interest retention. The recall candidate pool construction module calculates the user's current interest score on the corresponding topic based on the baseline interest score and interest retention rate; By setting a current interest threshold, the degree of user interest decay is determined; by setting a historical interest threshold, the intensity of user's historical interest is determined; by combining the results of the interest decay determination and the historical interest intensity determination, user-topic pairs that meet the preset conditions are selected to enter the recall candidate pool. The push sequence generation module generates recall candidate content based on user-topic pairs in the recall candidate pool. It prioritizes the recall candidate content by combining interest scores, forgetting level, and baseline interest scores, and generates a push sequence by combining user fatigue and push frequency limits. The memory content push module pushes memory content to users according to the push sequence.

Images