An agent-based full-process digital promotion method and system

Abstract

The application provides an agent-based full-process digital promotion method and system, relating to the technical field of business systems. The method comprises the following steps: capturing micro-behavior events, combining them into continuous event streams according to time sequence, performing slicing analysis, if it is detected that there is an intention label conflict pointing to different promotion targets within a preset short time window, it is determined that the intention is in a chaotic state, and the intention probability distribution of the target user for each candidate promotion target and its confidence score at the current time are output; if it is higher than a first threshold, a precise promotion strategy for the highest probability target is generated; if it is between the first threshold and a second threshold, a shunt comparison promotion strategy containing multiple comparison versions is generated; if it is lower than the second threshold, a bottom-line strategy of pushing general content is generated, and a channel fusing mechanism is triggered synchronously to lock high-frequency touch channels; the execution layer calls corresponding promotion channels for content delivery through API parallel calls according to the generated hierarchical strategy.

Description

Technical Field

[0001] This invention belongs to the field of commercial system technology, specifically relating to a full-process digital promotion method and system based on intelligent agents. Background Technology

[0002] In current digital marketing scenarios, users frequently generate rapid, continuous, and contradictory micro-behavioral sequences when interacting with promotional interfaces. For example, a user might click on promotional content for baby products and gaming peripherals within two seconds, and then return to the previous page. This conflict of intent labels within a short window (i.e., a state of intent chaos) is prevalent in real user behavior, especially in fragmented mobile internet usage scenarios.

[0003] Existing promotion systems typically employ two approaches: one is real-time decision-making based on single-point behavior, which promotes content based on the user's last click or the content with the longest dwell time; the other is session-level aggregation analysis, which waits for the user to complete the entire session before determining intent. The former can lead to promotion decisions being misled by accidental clicks or brief acts of curiosity, resulting in decision oscillations; the latter suffers from significant data lag, failing to respond in real time within the user's current session and missing promotion opportunities.

[0004] Specifically, if a user clicks on both baby and esports content within two seconds, and the system pushes an esports product based on the last click, the user's actual need might be for a baby product, rendering the push ineffective. If the system waits until the entire session ends before analyzing the content, the user may have already left. This "lag in intent recognition within micro-behavioral sequences" is a specific technical problem currently faced by real-time promotion systems, directly leading to decreased promotion accuracy and wasted budget. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the existing technology, a digital promotion method and system based on intelligent agents is provided.

[0006] Firstly, this application proposes a full-process digital promotion method based on intelligent agents, including the following steps:

[0007] By using a perception layer with real-time stream processing capabilities, the micro-behavioral events of target users in the promotional session interface are captured, each event is timestamped and combined into a continuous event stream according to the time sequence.

[0008] The temporal arbitration layer is invoked to perform slice analysis on the continuous event stream in a sliding window manner. If a conflict of intent tags pointing to different promotion targets is detected within a preset short time window, it is determined that the intent chaos state has been entered, and a lightweight attention model is triggered to perform weight calculation on the micro-behavior sequence within the window, and output the current time-time target user's intent probability distribution for each candidate promotion target and its confidence score.

[0009] The decision-making layer is triggered to generate a tiered strategy based on the preset threshold range where the confidence score falls. The preset threshold range includes a first threshold and a second threshold, and the first threshold is greater than the second threshold. If the confidence score is higher than the first threshold, a precise promotion strategy targeting the highest probability target is generated. If the confidence score is between the first threshold and the second threshold, a traffic diversion and comparison promotion strategy containing multiple comparison versions is generated. If the confidence score is lower than the second threshold, a minimum guarantee strategy for pushing general content is generated, and a channel circuit breaker mechanism is triggered simultaneously to lock high-frequency reach channels.

[0010] The execution layer calls the corresponding promotion channels in parallel through the API according to the generated hierarchical strategy to deliver content, and continuously captures the feedback behavior of the target users after delivery. The feedback behavior is fed back to the perception layer as new micro-behavioral event inputs for online updating of the parameters of the attention model.

[0011] as well as,

[0012] Extract the trajectory of the target user's intent state before and after each touch. The trajectory of intent state evolution is generated in real time by the micro-behavioral events captured by the perception layer and determined by the temporal arbitration layer. It includes the duration of intent chaos within the preset window before the touch, whether the intent state changes after the touch, and the duration of the clear intent state after the change.

[0013] The intent state evolution trajectory is used as the immediate impact feature of this outreach, and is stored together with the outreach time and policy version identifier in the outreach impact record table of the target user;

[0014] When a related behavior is captured, the reach impact record table of the target user is invoked to identify all historical reach events that occurred within the validity period and resulted in a positive change in intent state after the reach. The positive change is used as candidate evidence that the reach has a long-term impact. The validity period is a preset attribution decay window length, and the positive change is the change in user intent state from chaotic to clear.

[0015] According to the technical solution provided in this application, the generation of a traffic diversion and comparison promotion strategy containing multiple comparison versions includes the following steps:

[0016] In response to the decision layer generating a traffic-splitting and comparison promotion strategy that includes a first strategy version and a second strategy version, a unique session identifier for the target user is obtained;

[0017] Using the unique session identifier as input, a consistent hashing algorithm is used to map it to a preset traffic bucket space to determine the unique policy version bucket to which the target user belongs;

[0018] Based on the unique strategy version bucket, a cross-channel outreach instruction is generated for the target user. The cross-channel outreach instruction carries the unique strategy version identifier corresponding to the target user. The cross-channel outreach instruction is sent to the execution layer according to the preset channel priority order so that the target user only receives the promotional content corresponding to the unique strategy version identifier on all outreach channels.

[0019] The unique strategy version identifier is bound and stored with the subsequently captured feedback behavior of the target user, and used as an attribution basis when updating the attention model parameters online.

[0020] According to the technical solution provided in this application, the promotional content corresponding to the unique strategy version identifier is generated through the following steps:

[0021] Obtain the static profile data and historical behavior sequence of the target user, and construct the personalized feature vector of the target user by combining it with the real-time intent probability distribution of the current session;

[0022] The content generation model is invoked, with the unique strategy version identifier as a constraint, and the personalized feature vector is input to generate specific promotional content materials that conform to the tone of the strategy version. The specific promotional content materials are different among different users.

[0023] The generated promotional content materials are populated into the cross-channel outreach instruction, and then delivered by the execution layer through the corresponding channels;

[0024] In the feedback behavior captured after deployment, the unique strategy version identifier and the personalized feature vector are bound together to update the attention model parameters and content generation model parameters respectively.

[0025] According to the technical solution provided in this application, the preset traffic bucket space is dynamically updated through the following steps:

[0026] Using a preset time window as the period, the accumulated feedback behavior data of each strategy version within the time window is statistically analyzed, and the effect indicators and confidence intervals of each strategy version are calculated.

[0027] A traffic redistribution event is triggered when at least one strategy version has a significantly better performance metric than other versions and its confidence interval width is less than a preset convergence threshold.

[0028] Based on the performance metrics ratio of each strategy version, the hash value range corresponding to each strategy version in the traffic bucket space is recalculated so that the strategy version with better performance gets a larger range share, while maintaining the continuity and coverage integrity of the range of each strategy version.

[0029] The updated hash value range will be synchronized to the traffic distribution arbitration module so that new target users will be allocated to each policy version according to the new ratio, while the policy versions of previously allocated users will remain unchanged.

[0030] According to the technical solution provided in this application, after binding and storing the unique policy version identifier with the subsequently captured feedback behavior of the target user, the method further includes the following steps:

[0031] Acquire the instant feedback behavior generated by the target user within a preset short window after this contact. The instant feedback behavior is an interactive operation that can be directly attributed, including clicking, closing, or swiping.

[0032] Calculate the confidence level of the instant feedback reached this time based on the type and intensity of the instant feedback behavior;

[0033] If the confidence level of the instant feedback is higher than or equal to the preset instant confidence threshold, the instant feedback behavior will be used directly as the basis for evaluating the effectiveness of the strategy version, and the tracking process for this reach will end.

[0034] According to the technical solution provided in this application, after calculating the confidence level of the immediate feedback received this time, the method further includes the following steps:

[0035] If the confidence level of the instant feedback is lower than the preset instant confidence threshold, the target user's cross-session behavior tracking chain is triggered. The behavior tracking chain starts from the moment of this contact and has a preset decay window length as the validity period. It continuously captures all related behaviors generated by the user within the validity period. The related behaviors include delayed behaviors such as active search, deep browsing, adding to favorites and shopping cart, or cross-session conversion.

[0036] When a related behavior is captured, the attribution weight is dynamically calculated based on the behavior type and the time interval between the current contact. The shorter the time interval and the higher the relevance between the behavior and the contacted content, the greater the attribution weight.

[0037] Based on the attribution weights, calculate the weighted effect index of each strategy version within the decay window, and use the weighted effect index as input to update the attention model parameters and content generation model parameters;

[0038] The strategy version identifier reached this time is bound and stored with the weighted effect metric for subsequent dynamic updates of the traffic bucket space.

[0039] According to the technical solution provided in this application, before dynamically calculating the attribution weight based on the time interval between the behavior type and the current contact, the following steps are also included:

[0040] Extract the trajectory of the target user's intent state before and after each touch. The trajectory of intent state evolution is generated in real time by the micro-behavioral events captured by the perception layer and determined by the temporal arbitration layer. It includes the duration of intent chaos within the preset window before the touch, whether the intent state changes after the touch, and the duration of the clear intent state after the change.

[0041] The intent state evolution trajectory is used as the immediate impact feature of this outreach, and is stored together with the outreach time and policy version identifier in the outreach impact record table of the target user;

[0042] When a related behavior is captured, the reach impact record table of the target user is invoked to identify all historical reach events that occurred within the validity period and whose intent state underwent a positive transformation after the reach. The positive transformation is used as candidate evidence that the reach had a long-term impact.

[0043] According to the technical solution provided in this application, before dynamically calculating the attribution weight based on the time interval between the behavior type and the current contact, the following steps are also included:

[0044] Based on the aforementioned outreach impact record table, determine whether there are any historical outreach efforts that have been identified as generating positive conversions within the validity period;

[0045] The dynamic calculation of attribution weights based on the time interval between the behavior type and the current contact includes the following steps:

[0046] If it does not exist, the attribution weight is dynamically calculated based on the behavior type and the time interval between the current contact.

[0047] According to the technical solution provided in this application, after determining whether there are any historical contacts identified as generating positive conversions within the validity period, the method further includes the following steps:

[0048] All historical touches identified as generating positive transitions are acquired. For each historical touch, the time interval between its touch time and the time of the associated behavior is calculated, and the intensity of the intent state transition corresponding to that touch is extracted. The transition intensity is determined by the ratio of the duration of the intent state after the touch to the duration of the intent state chaos before the touch.

[0049] The time interval and the intensity of the intent state transition are input into a preset attribution weight calculation function to generate the contribution weight of this touch on the current associated behavior. The shorter the time interval and the greater the transition intensity, the higher the contribution weight.

[0050] Normalize the contribution weights of all historical outreaches, and then distribute the effects of related behaviors to the corresponding strategy versions of each outreach according to the normalized weights.

[0051] The allocated performance metrics are accumulated into the performance statistics of each strategy version for subsequent dynamic updates of traffic bucket space and optimization of attention model parameters.

[0052] Secondly, this application proposes a full-process digital promotion system based on intelligent agents, used to implement the full-process digital promotion method based on intelligent agents as described above, including:

[0053] A perceptual agent is configured to capture micro-behavioral events of a target user on a promotional session interface through a perception layer with real-time stream processing capabilities, timestamp each event, and combine them into a continuous event stream according to the time sequence.

[0054] The intelligent agent is configured to call the temporal arbitration layer to perform slice analysis on the continuous event stream in a sliding window manner. If a conflict of intent tags pointing to different promotion targets is detected within a preset short window, it is determined to enter the intent chaos state and triggers a lightweight attention model to perform weight calculation on the micro-behavior sequence within the window, and outputs the current time-time target user's intent probability distribution for each candidate promotion target and its confidence score.

[0055] A strategy generation agent is configured to trigger the decision layer to execute a tiered strategy generation based on a preset threshold range in which the confidence score falls: if it is higher than the first threshold, a precise promotion strategy targeting the highest probability target is generated; if it is between the first and second thresholds, a diversion and comparison promotion strategy containing multiple comparison versions is generated; if it is lower than the second threshold, a minimum guarantee strategy for pushing general content is generated, and a channel circuit breaker mechanism is triggered simultaneously to lock high-frequency reach channels.

[0056] An execution agent is configured to call the execution layer to call the corresponding promotion channels in parallel via API according to the generated hierarchical strategy for content delivery, and continuously capture the feedback behavior of the target user after delivery, and feed the feedback behavior back to the perception layer as new micro-behavioral event input for online updating of the parameters of the attention model;

[0057] as well as,

[0058] Extraction agent, the extraction agent being configured for:

[0059] Extract the trajectory of the target user's intent state before and after each touch. The trajectory of intent state evolution is generated in real time by the micro-behavioral events captured by the perception layer and determined by the temporal arbitration layer. It includes the duration of intent chaos within the preset window before the touch, whether the intent state changes after the touch, and the duration of the clear intent state after the change.

[0060] The intent state evolution trajectory is used as the immediate impact feature of this outreach, and is stored together with the outreach time and policy version identifier in the outreach impact record table of the target user;

[0061] When a related behavior is captured, the reach impact record table of the target user is invoked to identify all historical reach events that occurred within the validity period and resulted in a positive change in intent state after the reach. The positive change is used as candidate evidence that the reach has a long-term impact. The validity period is a preset attribution decay window length, and the positive change is the change in user intent state from chaotic to clear.

[0062] Compared with existing technologies, the advantages of this application are as follows: First, it achieves real-time chaotic recognition of micro-behavioral sequences: by capturing timestamped micro-behavioral event streams through the perception layer and combining them with sliding window analysis of the temporal arbitration layer, it can identify user intent conflicts generated in a very short time within milliseconds, solving the problem of misjudgment of intent caused by ignoring the temporal relationship of behaviors in traditional systems. Second, it introduces an attention mechanism to quantify intent probability: a lightweight attention model is used to calculate the weights of micro-behavioral sequences within the chaotic window, outputting the intent probability distribution and confidence score of each candidate promotion target. This allows the system to make quantifiable and reliable intent inferences even when user intent is unclear, rather than simply relying on the last click. Third, it generates hierarchical strategies based on confidence: according to different threshold ranges of the confidence score, it dynamically generates precise promotion strategies, traffic-dividing comparison promotion strategies, or safety net strategies, avoiding the resource waste caused by a one-size-fits-all approach when user intent is chaotic, and achieving adaptive matching between promotion strategies and the user's real-time intent state. It can also form a closed-loop optimization mechanism for the entire process: the feedback behavior after deployment is fed back to the perception layer as new micro-behavioral event input, which is used to update the attention model parameters online, so that the system can continuously adapt to changes in user behavior patterns and continuously improve the accuracy of intent recognition as data accumulates. Attached Figure Description

[0063] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0064] Figure 1 A flowchart illustrating the steps of the agent-based full-process digital promotion method provided in this application. Detailed Implementation

[0065] The present application will now be described in further 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 not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0066] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] Example 1

[0068] As mentioned in the background section regarding the technical problems, this application proposes a full-process digital promotion method based on intelligent agents, such as... Figure 1 As shown, it includes the following steps:

[0069] S1. Through a perception layer with real-time stream processing capabilities, capture the micro-behavioral events of target users in the promotional session interface, mark each event with a timestamp, and combine them into a continuous event stream according to the time sequence;

[0070] S2. Call the temporal arbitration layer to perform slice analysis on the continuous event stream in a sliding window manner. If a conflict of intent tags pointing to different promotion targets is detected within a preset short window, it is determined that the intent chaos state has been entered, and the lightweight attention model is triggered to perform weight calculation on the micro-behavior sequence within the window, and output the current time target user's intent probability distribution for each candidate promotion target and its confidence score.

[0071] S3. Trigger the decision layer to generate a tiered strategy based on the preset threshold range where the confidence score is located. The preset threshold range includes a first threshold and a second threshold, and the first threshold is greater than the second threshold. If it is higher than the first threshold, generate a precise promotion strategy for the highest probability target. If it is between the first threshold and the second threshold, generate a diversion and comparison promotion strategy containing multiple comparison versions. If it is lower than the second threshold, generate a minimum guarantee strategy for pushing general content, and simultaneously trigger a channel circuit breaker mechanism to lock high-frequency reach channels.

[0072] S4. The execution layer calls the corresponding promotion channels in parallel via API according to the generated hierarchical strategy to deliver content, and continuously captures the feedback behavior of the target users after delivery. The feedback behavior is fed back to the perception layer as new micro-behavioral event inputs for online updating of the parameters of the attention model.

[0073] First, a perception layer with real-time stream processing capabilities captures micro-behavioral events of target users on the promotional session interface. The perception layer is deployed at the data access point on the application front-end or server side, continuously monitoring various operations generated when the user interacts with the interface. These micro-behavioral events refer to extremely fine-grained units of user interaction on the interface, such as mouse movement, page scrolling speed and depth, focus gain and loss of input boxes, hover duration of controls, clicks, swipes, long presses, and even more subtle operations like taking screenshots or switching the page from the background to the foreground. The perception layer marks each captured micro-behavioral event with a precise timestamp and organizes these events into a continuous event stream strictly according to their natural chronological order, then transmits this event stream to the next processing layer in real time.

[0074] Subsequently, the temporal arbitration layer is invoked to slice and analyze the continuous event stream using a sliding window approach. Upon receiving the event stream, the temporal arbitration layer initiates a continuously sliding analysis window, for example, setting a short window of 3 seconds and sliding forward once per second. This layer performs semantic parsing on all micro-behavioral events falling within the current window, identifying the potential promotional target pointed to by each event, such as a specific product, a promotional activity, or a brand zone. If a conflict is detected within the preset short window between intent tags pointing to two or more different promotional targets—for example, if a user clicks on the details link of product A, closes the promotional pop-up of product A, and simultaneously quickly swipes past the promotional image of product B within 2 seconds—the temporal arbitration layer determines that the user has entered an intent chaos state. An intent chaos state refers to a user exhibiting contradictory or wavering behavior, without a clear decision-making tendency. Once this state is determined, the system immediately triggers a pre-trained lightweight attention model. This model calculates weights for the sequence of micro-behaviors within the window, automatically assigning higher attention weights to behaviors that better represent the user's final decision-making tendency. For example, the weight of the "add to cart" behavior is set higher than that of the "simply browse" behavior. After calculation, the model outputs the probability distribution of the target user's intent towards each candidate promotional target at the current moment. For example, the probability that the user will buy product A is 60%, the probability that they want to know about the discount on product B is 30%, and the probability that they have no clear intent is 10%. It also outputs the confidence score for this prediction, for example, 75%. The confidence score represents the model's degree of certainty about its own judgment.

[0075] Next, the decision-making layer is triggered to generate a tiered strategy based on the preset threshold range of the confidence score. The decision-making layer presets two key thresholds: a first threshold and a second threshold, with the first threshold being higher than the second. When the received confidence score is higher than the first threshold, for example, above 90%, the decision-making layer considers the system's judgment of the user's intent to be very certain. At this point, a precise promotion strategy targeting the highest probability objective is generated, such as directly pushing a limited-time exclusive coupon for the product to drive conversion. When the confidence score is between the first and second thresholds, for example, between 60% and 90%, the decision-making layer considers the user's intent to have some inclination but is not yet clear. At this point, a differentiated promotion strategy is generated, i.e., two or more different versions of content are prepared, such as one version emphasizing the product's cost-effectiveness parameters, and another version telling the product's brand story, so that a better solution can be determined through subsequent comparative testing. When the confidence score falls below the second threshold, such as below 60%, the decision-making level considers that the system is completely unable to grasp the user's intent, and the user may be browsing aimlessly. In this case, a backup strategy of pushing general content is generated, such as the platform's popular recommendation list or public service announcements, to avoid user resentment due to incorrect pushes. At the same time, the decision-making level simultaneously triggers the channel circuit breaker mechanism, temporarily locking those channels that frequently reach the user, such as SMS push or application notifications, suspending the push of any information to the user through these channels until subsequent behavior shows a clearer intent.

[0076] Finally, the execution layer, based on the generated tiered strategy, uses the application programming interface (API) to call the corresponding promotion channels in parallel to deliver content. The execution layer encapsulates interfaces for connecting with various promotion channels, such as in-app pop-ups, news feed ads, SMS gateways, and email service providers. Upon receiving the decision instruction, the execution layer accurately delivers the content to the target users by calling the corresponding channel's API. After delivery, the execution layer doesn't stop working but continuously captures the target user's feedback behavior after receiving the promotional content, such as whether they clicked, participated in activities, or ultimately completed a purchase conversion. These feedback behaviors are treated as new micro-behavioral events and flow back to the perception layer, forming a closed loop. The perception layer then timestamps these feedback behaviors again and inputs them into the event stream to update the parameters of the attention model online, enabling the model to self-optimize based on the latest delivery results, thereby achieving dynamic iteration and continuous evolution of the promotion strategy.

[0077] This embodiment constructs a four-layer intelligent agent architecture encompassing perception, arbitration, decision-making, and execution. This enables real-time capture of user micro-behaviors and precise arbitration of chaotic intent states. Furthermore, it performs hierarchical decision-making based on confidence levels, resolving the issues of delayed intent recognition and rigid strategies in traditional promotion methods. Through a closed-loop feedback mechanism, the system can continuously learn and optimize, effectively controlling the risk of user disruption while improving promotion effectiveness, thus achieving a balance between promotion efficiency and user experience.

[0078] Furthermore, before dynamically calculating the attribution weight based on the behavior type and the time interval between the current contact, the following steps are also included:

[0079] Extract the trajectory of the target user's intent state before and after each touch. The trajectory of intent state evolution is generated in real time by the micro-behavioral events captured by the perception layer and determined by the temporal arbitration layer. It includes the duration of intent chaos within the preset window before the touch, whether the intent state changes after the touch, and the duration of the clear intent state after the change.

[0080] The intent state evolution trajectory is used as the immediate impact feature of this outreach, and is stored together with the outreach time and policy version identifier in the outreach impact record table of the target user;

[0081] When a related behavior is captured, the reach impact record table of the target user is invoked to identify all historical reach events that occurred within the validity period and resulted in a positive change in intent state after the reach. The positive change is used as candidate evidence that the reach has a long-term impact. The validity period is a preset attribution decay window length, and the positive change is the change in user intent state from chaotic to clear.

[0082] Specifically, this embodiment provides a detailed implementation method for introducing intent state evolution information in the delayed feedback attribution process, aiming to more accurately determine the long-term impact of promotional outreach by analyzing changes in user intent before and after user engagement. Before or simultaneously with the system triggering the cross-session behavior tracking chain, an additional data recording step is performed. Specifically, the system continuously extracts the intent state evolution trajectory of the target user before and after each promotional outreach. The intent state evolution trajectory is serialized information generated in real-time by the temporal arbitration layer from micro-behavioral events captured by the perception layer; it completely records the change process of user intent over time. The trajectory specifically includes three key dimensions of information: First, the duration of intent confusion within the preset window before the touchpoint, i.e., the length of time the user remains in a state of undecided intent before receiving the promotional content, such as the user remaining in a state of intent confusion for 5 seconds before the touchpoint; second, whether the intent state changes after the touchpoint, i.e., whether the user's intent confusion ends after receiving the promotional content and clearly points to a specific target, such as changing from a state of confusion to a state of clear interest in product A; and third, the duration of the clear intent state after the change, i.e., the time the user maintains this clear intent, such as the user continuously focusing on product A for 10 seconds after the touchpoint.

[0083] After acquiring the intent state evolution trajectory, the system uses this trajectory as the immediate impact feature of that outreach, storing it along with the outreach time and policy version identifier in the target user's outreach impact record table. The outreach impact record table is a data table maintained for each user, used to store the impact information generated from each promotional outreach received by that user throughout history. For example, each record in the table can contain fields such as: outreach timestamp, policy version identifier, duration of chaos before outreach, whether a conversion occurred after outreach, and duration of intent after conversion. This table provides rich contextual information for subsequent analysis of the long-term effects of each outreach.

[0084] When the system detects a relevant action within a subsequent decay window—for example, a user completes a purchase 7 days later—it no longer calculates the attribution weight solely based on time intervals. Instead, it first accesses the target user's outreach impact record table. The system retrieves all historical outreaches that occurred within the validity period and meet specific conditions. The validity period refers to the time interval between the current relevant action and the current action not exceeding the decay window length. The specific conditions refer to a positive shift in intent state after the outreach; that is, the user transitions from a chaotic state to a definite state consistent with the current relevant action's goal. For example, if the current relevant action is purchasing product A, the system will retrieve all historical outreach records within the validity period where the user's intent explicitly points to product A after the outreach. The system identifies these outreaches as candidate evidence that may have a long-term impact on the current relevant action, providing a basis for subsequent refined attribution.

[0085] This embodiment incorporates the dynamic changes in user intent into attribution analysis by introducing an intent state evolution trajectory and establishing a reach impact record table. Its technical advantage lies in moving beyond reliance solely on temporal proximity to determine causality, and instead introducing evidence of changes in the user's psychological state. If a reach successfully guides a user from a state of confusion to a state of clear intent, then even if the reach occurred early, it may have a decisive impact on subsequent conversions. Recording this information lays the foundation for more scientific attribution weight calculations, enabling the system to more accurately identify truly effective promotional outreach.

[0086] In a preferred embodiment, generating a traffic splitting and comparison promotion strategy comprising multiple control versions includes the following steps:

[0087] In response to the decision layer generating a traffic-splitting and comparison promotion strategy that includes a first strategy version and a second strategy version, a unique session identifier for the target user is obtained;

[0088] Using the unique session identifier as input, a consistent hashing algorithm is used to map it to a preset traffic bucket space to determine the unique policy version bucket to which the target user belongs;

[0089] Based on the unique strategy version bucket, a cross-channel outreach instruction is generated for the target user. The cross-channel outreach instruction carries the unique strategy version identifier corresponding to the target user. The cross-channel outreach instruction is sent to the execution layer according to the preset channel priority order so that the target user only receives the promotional content corresponding to the unique strategy version identifier on all outreach channels.

[0090] The unique strategy version identifier is bound and stored with the subsequently captured feedback behavior of the target user, and used as an attribution basis when updating the attention model parameters online.

[0091] Specifically, when the decision-making level determines, based on the arbitration result, that a traffic diversion and comparative promotion strategy needs to be implemented, the system first responds to this instruction and obtains the unique session identifier of the current target user. This unique session identifier is a globally unique credential used to distinguish different users; it can be the user's account ID after login, or a device ID generated by the system and stored in a browser cookie or mobile device advertising identifier when not logged in. The purpose of obtaining this identifier is to provide stable and unique identification of users.

[0092] Next, the system uses this unique session identifier as input and maps it to a preset traffic bucket space using a consistent hashing algorithm, thereby determining the unique policy version bucket to which the target user belongs. The traffic bucket space is an abstract address space, typically represented as a ring structure with its values ​​pre-divided into several consecutive intervals, each interval representing a specific policy version bucket. For example, version A corresponds to intervals 0 to 1000 on the ring, and version B corresponds to intervals 1001 to 2000. The characteristic of the consistent hashing algorithm is that for the same input identifier, after hash calculation, it will always fall into the same position on the ring, thus ensuring that the same user is always mapped to the same policy version bucket. Compared to ordinary modulo hashing, this algorithm can maximize the preservation of existing users' bucket affiliations and maintain the stability of test data when adjusting the traffic ratio of each version.

[0093] After determining the unique policy version bucket to which a user belongs based on the mapping results, the decision layer generates a cross-channel outreach instruction for that target user. This instruction carries a unique policy version identifier for the target user, such as version_id A or B. It also includes a preset channel priority order, for example, prioritizing outreach via in-app pop-ups, followed by feed recommendations, and finally SMS notifications. After generating the instruction, the system sends it to the execution layer according to the preset channel priority order. Upon receiving the instruction, the execution layer ensures that subsequent interactions with the user through various channels use promotional content corresponding to the unique policy version identifier carried in the instruction. This ensures that regardless of which channel a user encounters the promotional information, the content they see belongs to the same policy version, thus avoiding test data contamination and user experience confusion caused by users receiving different versions of content on different channels.

[0094] Finally, the system binds and stores the unique strategy version identifier used in this outreach with the subsequently captured feedback behavior of the target user. This binding storage means that a dedicated field is set up in the data structure recording user clicks, browsing, purchases, and other feedback behaviors to record the strategy version identifier corresponding to the promotional content that prompted that behavior. This version-identified feedback data will serve as an important attribution basis for subsequent model optimization. For example, when updating attention model parameters online, the system can use this data to determine which version of the promotional strategy is more likely to elicit positive feedback for user groups with certain behavioral characteristics, thereby adjusting the model's preference weights for different strategies, making future intent recognition and strategy selection more accurate.

[0095] This embodiment introduces a consistent hashing algorithm and a unique strategy version identifier binding mechanism to provide a scientific traffic splitting and attribution scheme for comparative promotion strategies. Its technical effect is to ensure that each user only encounters a single version of the promotional content during the testing period, guaranteeing the purity of data from parallel testing of multiple versions and avoiding cross-interference. Simultaneously, by forcibly binding the version identifier with user feedback behavior, it achieves version-level accurate effect attribution, providing a high-quality decision-making basis for subsequent model optimization and traffic allocation, thereby significantly improving the scientific rigor and credibility of A / B testing.

[0096] In a preferred embodiment, the promotional content corresponding to the unique strategy version identifier is generated through the following steps:

[0097] Obtain the static profile data and historical behavior sequence of the target user, and construct the personalized feature vector of the target user by combining it with the real-time intent probability distribution of the current session;

[0098] The content generation model is invoked, with the unique strategy version identifier as a constraint, and the personalized feature vector is input to generate specific promotional content materials that conform to the tone of the strategy version. The specific promotional content materials are different among different users.

[0099] The generated promotional content materials are populated into the cross-channel outreach instruction, and then delivered by the execution layer through the corresponding channels;

[0100] In the feedback behavior captured after deployment, the unique strategy version identifier and the personalized feature vector are bound together to update the attention model parameters and content generation model parameters respectively.

[0101] Specifically, this embodiment further refines the specific generation method of promotional content corresponding to the unique strategy version identifier, aiming to achieve deep personalization within the strategy framework. Before or simultaneously with determining the unique strategy version bucket to which the target user belongs and generating cross-channel outreach instructions, the system begins preparing to generate highly personalized promotional content for that user. First, the system acquires the target user's static profile data and historical behavior sequence. Static profile data includes demographic information filled in by the user during registration or accumulated by the system over a long period, such as age, gender, occupation, region, and stable interest tags summarized from long-term behavior. The historical behavior sequence is the user's behavior logs extracted from the data warehouse over a period of time, such as browsing history, click history, search keywords, favorites, and purchased items over the past 7 or 30 days. This data reflects the user's long-term and medium-term interest preferences. At the same time, the system also combines the real-time intent probability distribution of the current session, which has just been calculated by the temporal arbitration layer and represents the user's current interest in various promotional targets. Subsequently, the system fuses and vectorizes these three parts of information to construct the personalized feature vector of the target user at the current moment. This vector fully encapsulates the user's identity attributes, historical preferences, and real-time intent.

[0102] Next, the system invokes a pre-trained content generation model. This model can be a deep learning-based generative model, such as the GPT series models for text generation, or a cross-modal generative model for image and text generation. When invoking the model, the system uses a unique strategy version identifier as a constraint on the generation process, while simultaneously inputting the personalized feature vector constructed in the previous step. The strategy version identifier, as a constraint, determines the macro-direction and tone of content generation. For example, if strategy version A is defined as "emphasizing functional parameters and cost-effectiveness," the model will be guided to explore and highlight elements such as the product's price advantages, performance parameters, and comparative data with competitors when generating content; if strategy version B is defined as "emphasizing emotional resonance and lifestyle," the model will be guided to connect with the user's life scenarios, tell the brand story, and evoke emotional identification. Under the constraint of the strategy version identifier, the model, combined with the personalized feature vector, ultimately generates specific promotional content materials that conform to the tone of the strategy version and fit the specific characteristics of the user. Because each user has a different personalized feature vector, even if they fall into the same strategy version bucket, the promotional content that different users ultimately see will vary significantly in details. For example, the focus of the copy, the style of the images, and the specific product models recommended may all vary from person to person.

[0103] After generating specific promotional content materials, the system populates them into cross-channel outreach instructions, which are then distributed through the appropriate channels by the execution layer. User feedback captured after distribution, in addition to being bound to a unique strategy version identifier, is also bound to a personalized feature vector used to generate the content. This means that each user feedback simultaneously records "who saw which version of the content" and "what user characteristics generated this content." This high-value data is used to update two models. On one hand, the combination of feedback behavior and strategy version identifier is used to update the attention model parameters, helping the model more accurately learn the applicability of different strategy versions in different contexts. On the other hand, the combination of feedback behavior and personalized feature vectors is used to update the parameters of the content generation model itself, enabling the model to continuously optimize through reinforcement learning and other methods, learning which specific content elements, under the constraints of a specific strategy version, will most likely elicit positive feedback from user groups with specific feature vectors. This mechanism of parallel optimization of two models constitutes a more powerful learning loop.

[0104] This embodiment combines strategy version constraints with personalized feature vectors, utilizing a content generation model to dynamically generate promotional content, achieving deep personalization within a traffic-splitting and comparative testing framework. Its technical advantage lies in breaking the limitations of traditional A / B testing's one-size-fits-all approach, allowing each test version to be dynamically varied, thus enabling the exploration of more refined user preferences. Simultaneously, feedback data is used to update both the intent recognition model and the content generation model, constructing a dual-drive optimization flywheel. The more accurate the intent recognition, the more relevant the content generation; the more relevant the content, the more valuable the feedback data. The feedback data, in turn, optimizes both models simultaneously, greatly improving the system's evolution speed and personalization capabilities, ultimately maximizing the promotional effect.

[0105] In a preferred embodiment, the preset flow bucket space is dynamically updated through the following steps:

[0106] Using a preset time window as the period, the accumulated feedback behavior data of each strategy version within the time window is statistically analyzed, and the effect indicators and confidence intervals of each strategy version are calculated.

[0107] A traffic redistribution event is triggered when at least one strategy version has a significantly better performance metric than other versions and its confidence interval width is less than a preset convergence threshold.

[0108] Based on the performance metrics ratio of each strategy version, the hash value range corresponding to each strategy version in the traffic bucket space is recalculated so that the strategy version with better performance gets a larger range share, while maintaining the continuity and coverage integrity of the range of each strategy version.

[0109] The updated hash value range will be synchronized to the traffic distribution arbitration module so that new target users will be allocated to each policy version according to the new ratio, while the policy versions of previously allocated users will remain unchanged.

[0110] This embodiment provides a specific implementation method for dynamically updating the preset traffic bucket space, aiming to achieve real-time optimization of traffic ratios during multi-version comparison testing, allowing the more effective strategy version to receive more exposure opportunities. The dynamic update of the traffic bucket space is performed periodically within a preset time window. The system sets a fixed time interval, such as every hour or every 24 hours, as an evaluation cycle. At the end of each cycle, the system automatically collects the accumulated feedback behavior data for each strategy version within that time window. Feedback behavior data includes, but is not limited to, key indicators reflecting promotional effectiveness such as user clicks, conversions, dwell time, and add-to-cart behavior. Based on these statistical data, the system calculates the performance indicators and their confidence intervals for each strategy version. Performance indicators can be click-through rate, conversion rate, or a comprehensive weighted score, while the confidence interval reflects the statistical reliability of the performance indicator; a narrower interval indicates more stable and reliable statistical results. When the system detects that at least one strategy version's performance indicator is significantly better than other versions, and the confidence interval width of that version is less than a preset convergence threshold, a traffic redistribution event is triggered. "Significantly better" here means that statistical testing confirms that the performance difference between versions is statistically significant, rather than caused by random fluctuations. The preset convergence threshold is used to determine whether the statistical results are sufficiently stable. For example, setting the confidence interval width to no more than 10% of the performance index value is a convergence condition. Only when both conditions are met simultaneously does the system consider there sufficient evidence that a certain version is indeed superior, and traffic adjustments can be made.

[0111] After a traffic redistribution event is triggered, the system recalculates the hash value range for each strategy version within the traffic bucket space based on the performance metrics ratios of each strategy version. The calculation principle is to allocate a larger share of the range to the strategy version with better performance, thereby allowing more new users to fall into that version. For example, if version A and version B originally each accounted for 50% of the traffic, and after a period of testing, version A's conversion rate is 5% and version B's conversion rate is 3%, and the statistical results are stable, then the system can readjust the ranges, expanding version A's share to 70% and reducing version B's share to 30%. During the adjustment process, the system must maintain the continuity and coverage integrity of the ranges for each strategy version, ensuring that every point on the entire hash ring uniquely belongs to a specific strategy version, without overlap or gaps.

[0112] Finally, the updated hash value range is synchronized to the traffic distribution arbitration module. The traffic distribution arbitration module is responsible for performing consistent hash mapping. After synchronization, new target users entering the system will be allocated to various policy versions according to the updated proportions. It's important to note that previously allocated users—those who completed hash mapping and received promotional content in the previous cycle—will retain their original policy version, which will not change due to adjustments in traffic ratios. This design ensures consistency in the promotional strategy for the same user, avoiding data pollution and a decline in user experience caused by mid-cycle version switching.

[0113] This embodiment achieves automated optimization of traffic bucket space by introducing periodic statistical evaluation and dynamic interval adjustment mechanisms. This transforms multi-version comparison testing from a static, proportional traffic allocation process into a dynamic adjustment based on real-time feedback data. It automatically distributes more exposure opportunities to the better-performing strategy version, thereby continuously improving the overall promotional effect during the testing process. Simultaneously, by maintaining the historical user versions unchanged, it balances the scientific nature of the testing with the stability of the user experience.

[0114] In a preferred embodiment, after binding and storing the unique policy version identifier with the subsequently captured feedback behavior of the target user, the method further includes the following steps:

[0115] Acquire the instant feedback behavior generated by the target user within a preset short window after this contact. The instant feedback behavior is an interactive operation that can be directly attributed, including clicking, closing, or swiping.

[0116] Calculate the confidence level of the instant feedback reached this time based on the type and intensity of the instant feedback behavior;

[0117] If the confidence level of the instant feedback is higher than or equal to the preset instant confidence threshold, the instant feedback behavior will be used directly as the basis for evaluating the effectiveness of the strategy version, and the tracking process for this reach will end.

[0118] Specifically, this embodiment provides a detailed implementation method for the feedback processing after the implementation of the diversion and comparison promotion strategy, especially for the capture and attribution mechanism of short-term immediate feedback.

[0119] After the system delivers promotional content carrying a unique policy version identifier to the target user, it continuously captures the user's immediate feedback behavior within a preset short window after the initial contact. The preset short window is a relatively short time period, such as within 30 seconds or 5 minutes after delivery, used to capture the user's direct reaction to the promotional content. Immediate feedback behavior refers to interactive operations that can be directly attributed to the initial contact and whose relevance can be determined without complex attribution logic. These include actions such as clicking on promotional content, closing promotional pop-ups, quickly swiping on the content page, or directly returning. These behaviors occur very close to the initial contact time, with a clear causal relationship. After capturing immediate feedback behavior, the system calculates the confidence level of the immediate feedback based on the type and intensity of these behaviors. Immediate feedback confidence level is a quantitative indicator used to assess the extent to which the feedback behavior can be identified as a direct reaction to the promotional content. Different types of immediate feedback behavior are assigned different basic weights; for example, clicking has the highest weight and can be considered strong positive feedback; swiping has a medium weight and can be considered general attention; and directly closing has a low weight and can be considered negative feedback. Simultaneously, the response time of the behavior also affects the confidence score; the shorter the response time, the higher the confidence score. The system comprehensively calculates an integrated instant feedback confidence score, such as a value between 0 and 1, by combining the behavior type weight and the response time factor. After calculating the instant feedback confidence score, the system compares it with a preset instant confidence threshold. The preset instant confidence threshold is a pre-defined critical value, such as 0.8, used to determine whether the instant feedback is sufficiently reliable as a basis for the final effect evaluation. If the instant feedback confidence score is higher than or equal to this threshold, the system considers the attribution relationship of the feedback behavior to be very clear, and there is no need for complex delay tracking. At this time, the system directly uses the instant feedback behavior captured this time as the effect evaluation basis for this strategy version and records it in the corresponding effect statistics database. After the recording is completed, the tracking process for this reach ends, and the system will not conduct further long-term tracking of this reach to save computing resources.

[0120] Furthermore, after calculating the confidence level of the immediate feedback received in this instance, the process further includes the following steps:

[0121] If the confidence level of the instant feedback is lower than the preset instant confidence threshold, the target user's cross-session behavior tracking chain is triggered. The behavior tracking chain starts from the moment of this contact and has a preset decay window length as the validity period. It continuously captures all related behaviors generated by the user within the validity period. The related behaviors include delayed behaviors such as active search, deep browsing, adding to favorites and shopping cart, or cross-session conversion.

[0122] When a related behavior is captured, the attribution weight is dynamically calculated based on the behavior type and the time interval between the current contact. The shorter the time interval and the higher the relevance between the behavior and the contacted content, the greater the attribution weight.

[0123] Based on the attribution weights, calculate the weighted effect index of each strategy version within the decay window, and use the weighted effect index as input to update the attention model parameters and content generation model parameters;

[0124] The strategy version identifier reached this time is bound and stored with the weighted effect metric for subsequent dynamic updates of the traffic bucket space.

[0125] Specifically, when the system detects that the confidence level is below a preset instant confidence threshold, such as below 0.8, the system believes that it cannot accurately assess the effectiveness of this outreach based solely on immediate behavior within a short window, because the user may need more time to think and make decisions, and their feedback behavior may be delayed. At this point, the system triggers a cross-session behavior tracking chain for the target user. The behavior tracking chain starts at the moment of this outreach and has a preset decay window length as its validity period, continuously capturing all related behaviors generated by the user within the validity period. The decay window length is a relatively long time period, such as 7 or 14 days, to cover potential delayed feedback from the user. Related behaviors include not only instant clicks but also deep behaviors that may take longer to occur, specifically including subsequent proactive search behavior, in-depth browsing of related content, adding products to favorites or shopping carts, and conversion behavior finally completed in another session. Although these behaviors occur with a time lag, they have a potential causal relationship with this outreach. When the system captures any related behavior within the decay window, it immediately triggers the attribution weight calculation process. The system dynamically calculates the attribution weight of each interaction based on the time interval between the specific type of behavior and the moment of the interaction. The calculation principle is that the shorter the time interval, the stronger the correlation between the behavior and the interaction, and the greater the attribution weight. Simultaneously, the higher the relevance between the behavior type and the content of the interaction, the greater the attribution weight. For example, if a user searches for keywords related to the promoted product within one hour of the interaction, their attribution weight will be much higher than if they accidentally visit the brand's homepage seven days later. The system calculates the weight using a preset decay function, which typically exhibits exponential decay over time, while also incorporating a weighted adjustment based on the relevance coefficient of the behavior type. After calculating the attribution weight, the system calculates the weighted effect metric for each strategy version within the decay window. For example, if a purchase conversion is valued at 1 unit and the attribution weight is calculated to be 0.6, then this conversion will be counted in the performance statistics of the corresponding strategy version with a weight of 0.6, instead of being directly counted as 1. These weighted effect metrics will serve as important inputs for subsequent updates to the attention model parameters and content generation model parameters, ensuring that model optimization is based on more accurate attribution results. Finally, the system binds and stores the policy version identifier of this reach with the calculated weighted effect index for subsequent dynamic updates of traffic bucket space, so that traffic allocation can be adjusted based on more accurate latency effect data.

[0126] This embodiment addresses the challenge of performance attribution in delayed conversion scenarios by introducing cross-session behavior tracking chains and dynamic attribution weight calculation. Its technical advantage lies in its ability to incorporate long-term effects—those undeniably triggered by promotional outreach but whose impact cannot be determined through immediate feedback—with appropriate weights into the statistical system, resulting in more comprehensive and accurate performance evaluation. The design of dynamically decaying weights acknowledges the causal relationship of delayed behavior while avoiding erroneous attribution of overly distant, incidental behaviors, thus improving the scientific rigor of attribution. Ultimately, this more accurate attribution data provides a more reliable basis for attention model optimization and traffic allocation, enabling the system to better capture and utilize users' long-term behavioral patterns.

[0127] In a preferred embodiment, before dynamically calculating the attribution weight based on the behavior type and the time interval between the current contact, the method further includes the following steps:

[0128] Based on the aforementioned outreach impact record table, determine whether there are any historical outreach efforts that have been identified as generating positive conversions within the validity period;

[0129] The dynamic calculation of attribution weights based on the time interval between the behavior type and the current contact includes the following steps:

[0130] If it does not exist, the attribution weight is dynamically calculated based on the behavior type and the time interval between the current contact.

[0131] This embodiment provides a specific implementation of the judgment logic before attribution weight calculation, aiming to adopt different attribution strategies based on the existence of positive conversion history. After the system captures the associated behavior and calls the target user's reach impact record table, before performing the actual attribution weight calculation, the system first performs a key judgment step. Based on the data in the reach impact record table, the system determines whether there are any historical reaches that have been identified as generating positive conversion within the validity period. The validity period here refers to all historical reaches that occurred before the current associated behavior and whose time interval between the current associated behavior and the current associated behavior does not exceed the preset decay window length. The judgment criteria for generating positive conversion are the same as described above, that is, after the reach, the user's intent state changes from a chaotic state to a clear state consistent with the goal of the current associated behavior.

[0132] Depending on the judgment result, the system enters different attribution weight calculation paths. If the judgment result is that there are no historical outreaches that have generated positive conversions—that is, although the user has received multiple promotions within the validity period, none of them have successfully guided the user from a state of confusion to a state of clear intent—then the system considers the occurrence of the current associated behavior to be mainly attributable to the cumulative effect of recent outreaches or the user's autonomous decision-making. In this case, the system directly calculates the attribution weight dynamically based on the time interval between the behavior type and the current outreach, according to the method described above. The outreaches involved in the calculation at this time are usually the one or a few outreaches most recent to the current associated behavior, and their weights are assigned according to a time decay function.

[0133] This embodiment introduces a pre-processing logic for attribution weight calculation by defining the existence of positive conversion reach. The technical advantage is that it allows the attribution process to distinguish between two different scenarios: a standard case where there is no obvious intent to change evidence, in which standard time-decay attribution is used; and a special case where there is intent to change evidence, requiring more complex attribution logic. This design avoids excessive consumption of computational resources in simple scenarios while ensuring that important causal evidence is not overlooked in complex scenarios, making the attribution system both efficient and accurate. The introduction of this judgment logic reflects the system's deep understanding of the user's decision-making process; a successful promotion is not only about triggering an immediate click, but also about changing the user's intention state, and this change itself is strong evidence of long-term effects.

[0134] In a preferred embodiment, after determining whether there are any historical contacts identified as generating a positive conversion within the validity period, the method further includes the following steps:

[0135] All historical touches identified as generating positive transitions are acquired. For each historical touch, the time interval between its touch time and the time of the associated behavior is calculated, and the intensity of the intent state transition corresponding to that touch is extracted. The transition intensity is determined by the ratio of the duration of the intent state after the touch to the duration of the intent state chaos before the touch.

[0136] The time interval and the intensity of the intent state transition are input into a preset attribution weight calculation function to generate the contribution weight of this touch on the current associated behavior. The shorter the time interval and the greater the transition intensity, the higher the contribution weight.

[0137] Normalize the contribution weights of all historical outreaches, and then distribute the effects of related behaviors to the corresponding strategy versions of each outreach according to the normalized weights.

[0138] The allocated performance metrics are accumulated into the performance statistics of each strategy version for subsequent dynamic updates of traffic bucket space and optimization of attention model parameters.

[0139] Specifically, this embodiment provides a detailed implementation method for attribution weight calculation when there is a history of positive conversion, aiming to achieve refined attribution allocation for multiple potentially influential outreaches. When the system determines that there is a historical outreach identified as having generated a positive conversion within the validity period, it enters the refined attribution process described in this embodiment. First, the system retrieves all historical outreach identified as having generated a positive conversion from the outreach impact record table. There may be more than one of these outreaches; for example, a user may have received three promotions within two weeks, and each time the user demonstrated a clear intent state after the outreach.

[0140] For each historical touchpoint acquired, the system extracts and calculates two key pieces of information. The first is the time interval between the touchpoint's moment and the current associated behavior's moment; a shorter interval generally indicates a stronger association. The second is the intensity of the intent state transition corresponding to that touchpoint. Transition intensity is a quantitative indicator determined by the ratio of the duration of the clear intent state after the touchpoint to the duration of the intent confusion before the touchpoint. For example, if the user was confused for 4 seconds before the touchpoint and the clear state lasted for 8 seconds after the touchpoint, the transition intensity is 2; if the user was confused for 10 seconds before the touchpoint and the clear state lasted only 2 seconds after the touchpoint, the transition intensity is 0.2. A higher transition intensity indicates a stronger and more lasting guiding effect of the touchpoint on the user's intent.

[0141] Subsequently, the system inputs the time interval and intent state transition strength of each historical outreach into a pre-defined attribution weight calculation function. The function is designed based on the principle that shorter time intervals and stronger transition strengths result in higher contribution weights. The function can take the form of a weighted product, such as weights equal to the transition strength divided by the logarithm of the time interval, or it can be a more complex machine learning model. Through this function, the system generates a contribution weight score for each historical outreach, reflecting the importance of that outreach in facilitating the final conversion.

[0142] After generating the contribution weights for all historical outreaches, the system normalizes these weights. Normalization aims to adjust the sum of all weights to 1, making them a proportional coefficient that can distribute the overall effect. For example, if the original contribution weights for three outreaches were 0.8, 0.5, and 0.2, they might become 0.53, 0.33, and 0.13 after normalization. Subsequently, the system allocates the effect of the current associated behavior to the corresponding strategy version for each outreach according to the normalized weights. If the current associated behavior is a purchase worth 100 yuan, then the strategy versions corresponding to the three outreaches will receive effect contributions of 53 yuan, 33 yuan, and 13 yuan respectively.

[0143] After allocation, the system will accumulate the allocated performance metrics into the performance statistics of each strategy version. This accumulated data will be used for two important purposes: first, for the dynamic updating of traffic bucket space, enabling traffic allocation to be adjusted based on more accurate long-term attribution results; second, for the optimization of attention model parameters, allowing the model to learn that those outreach strategies that can trigger high conversion intensity deserve greater attention in future intent recognition, even if the conversion time is far away.

[0144] This embodiment constructs a refined attribution model with multiple touchpoints and a long-term perspective by introducing intent state transition strength and comprehensive attribution weight calculation. Its technical advantage lies in its ability to scientifically handle complex scenarios where multiple historical touchpoints jointly influence a final conversion. It no longer simply attributes credit to the last click, but allocates the effect based on the actual contribution of each touchpoint to changing user intent. In particular, incorporating intent state transition strength into the weight calculation ensures that key touchpoints, while not directly leading to immediate conversion, successfully reverse the user's hesitation and receive due recognition. This attribution approach more closely reflects real user decision-making psychology, providing higher-quality decision-making basis for model optimization and traffic allocation, ultimately driving the overall promotion strategy towards a direction that more profoundly influences user intent.

[0145] Example 2

[0146] This embodiment proposes a full-process digital promotion system based on intelligent agents, used to implement the full-process digital promotion method based on intelligent agents as described in Embodiment 1, including:

[0147] A perceptual agent is configured to capture micro-behavioral events of a target user on a promotional session interface through a perception layer with real-time stream processing capabilities, timestamp each event, and combine them into a continuous event stream according to the time sequence.

[0148] The intelligent agent is configured to call the temporal arbitration layer to perform slice analysis on the continuous event stream in a sliding window manner. If a conflict of intent tags pointing to different promotion targets is detected within a preset short window, it is determined to enter the intent chaos state and triggers a lightweight attention model to perform weight calculation on the micro-behavior sequence within the window, and outputs the current time-time target user's intent probability distribution for each candidate promotion target and its confidence score.

[0149] A strategy generation agent is configured to trigger the decision layer to execute a tiered strategy generation based on a preset threshold range in which the confidence score falls: if it is higher than the first threshold, a precise promotion strategy targeting the highest probability target is generated; if it is between the first and second thresholds, a diversion and comparison promotion strategy containing multiple comparison versions is generated; if it is lower than the second threshold, a minimum guarantee strategy for pushing general content is generated, and a channel circuit breaker mechanism is triggered simultaneously to lock high-frequency reach channels.

[0150] An execution agent is configured to call the execution layer to make parallel calls to the corresponding promotion channels through APIs according to the generated hierarchical strategy for content delivery, and continuously capture the feedback behavior of the target users after delivery, and feed the feedback behavior back to the perception layer as new micro-behavioral event inputs for online updating of the parameters of the attention model.

[0151] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A full-process digital promotion method based on intelligent agents, characterized in that, Includes the following steps: By using a perception layer with real-time stream processing capabilities, the micro-behavioral events of target users in the promotional session interface are captured, each event is timestamped and combined into a continuous event stream according to the time sequence. The temporal arbitration layer is invoked to perform slice analysis on the continuous event stream in a sliding window manner. If a conflict of intent tags pointing to different promotion targets is detected within a preset short time window, it is determined that the intent chaos state has been entered, and a lightweight attention model is triggered to perform weight calculation on the micro-behavior sequence within the window, and output the current time-time target user's intent probability distribution for each candidate promotion target and its confidence score. The decision-making layer is triggered to generate a tiered strategy based on the preset threshold range where the confidence score falls. The preset threshold range includes a first threshold and a second threshold, and the first threshold is greater than the second threshold. If the confidence score is higher than the first threshold, a precise promotion strategy targeting the highest probability target is generated. If the confidence score is between the first threshold and the second threshold, a traffic diversion and comparison promotion strategy containing multiple comparison versions is generated. If the confidence score is lower than the second threshold, a minimum guarantee strategy for pushing general content is generated, and a channel circuit breaker mechanism is triggered simultaneously to lock high-frequency reach channels. The execution layer calls the corresponding promotion channels in parallel through the API according to the generated hierarchical strategy to deliver content, and continuously captures the feedback behavior of the target users after delivery. The feedback behavior is fed back to the perception layer as new micro-behavioral event inputs for online updating of the parameters of the attention model. as well as, Extract the trajectory of the target user's intent state before and after each touch. The trajectory of intent state evolution is generated in real time by the micro-behavioral events captured by the perception layer and determined by the temporal arbitration layer. It includes the duration of intent chaos within the preset window before the touch, whether the intent state changes after the touch, and the duration of the clear intent state after the change. The intent state evolution trajectory is used as the immediate impact feature of this outreach, and is stored together with the outreach time and policy version identifier in the outreach impact record table of the target user; When a related behavior is captured, the reach impact record table of the target user is invoked to identify all historical reach events that occurred within the validity period and resulted in a positive change in intent state after the reach. The positive change is used as candidate evidence that the reach has a long-term impact. The validity period is a preset attribution decay window length, and the positive change is the change in user intent state from chaotic to clear.

2. The full-process digital promotion method based on intelligent agents according to claim 1, characterized in that, The generation of a traffic splitting and comparison promotion strategy containing multiple comparison versions includes the following steps: In response to the decision layer generating a traffic-splitting and comparison promotion strategy that includes a first strategy version and a second strategy version, a unique session identifier for the target user is obtained; Using the unique session identifier as input, a consistent hashing algorithm is used to map it to a preset traffic bucket space to determine the unique policy version bucket to which the target user belongs; Based on the unique strategy version bucket, a cross-channel outreach instruction is generated for the target user. The cross-channel outreach instruction carries the unique strategy version identifier corresponding to the target user. The cross-channel outreach instruction is sent to the execution layer according to the preset channel priority order so that the target user only receives the promotional content corresponding to the unique strategy version identifier on all outreach channels. The unique strategy version identifier is bound and stored with the subsequently captured feedback behavior of the target user, and used as an attribution basis when updating the attention model parameters online.

3. The full-process digital promotion method based on intelligent agents according to claim 2, characterized in that, The promotional content corresponding to the unique strategy version identifier is generated through the following steps: Obtain the static profile data and historical behavior sequence of the target user, and construct the personalized feature vector of the target user by combining it with the real-time intent probability distribution of the current session; The content generation model is invoked, with the unique strategy version identifier as a constraint, and the personalized feature vector is input to generate specific promotional content materials that conform to the tone of the strategy version. The specific promotional content materials are different among different users. The generated promotional content materials are populated into the cross-channel outreach instruction, and then delivered by the execution layer through the corresponding channels; In the feedback behavior captured after deployment, the unique strategy version identifier and the personalized feature vector are bound together to update the attention model parameters and content generation model parameters respectively.

4. The full-process digital promotion method based on intelligent agents according to claim 2, characterized in that, The preset flow bucket space is dynamically updated through the following steps: Using a preset time window as the period, the accumulated feedback behavior data of each strategy version within the time window is statistically analyzed, and the effect indicators and confidence intervals of each strategy version are calculated. A traffic redistribution event is triggered when at least one strategy version has a significantly better performance metric than other versions and its confidence interval width is less than a preset convergence threshold. Based on the performance metrics ratio of each strategy version, the hash value range corresponding to each strategy version in the traffic bucket space is recalculated so that the strategy version with better performance gets a larger range share, while maintaining the continuity and coverage integrity of the range of each strategy version. The updated hash value range will be synchronized to the traffic distribution arbitration module so that new target users will be allocated to each policy version according to the new ratio, while the policy versions of previously allocated users will remain unchanged.

5. The full-process digital promotion method based on intelligent agents according to claim 2, characterized in that, After binding and storing the unique policy version identifier with the subsequently captured feedback behavior of the target user, the method further includes the following steps: Acquire the instant feedback behavior generated by the target user within a preset short window after this contact. The instant feedback behavior is an interactive operation that can be directly attributed, including clicking, closing, or swiping. Calculate the confidence level of the instant feedback reached this time based on the type and intensity of the instant feedback behavior; If the confidence level of the instant feedback is higher than or equal to the preset instant confidence threshold, the instant feedback behavior will be used directly as the basis for evaluating the effectiveness of the strategy version, and the tracking process for this reach will end.

6. The full-process digital promotion method based on intelligent agents according to claim 5, characterized in that, After calculating the confidence level of the immediate feedback received this time, the following steps are also included: If the confidence level of the instant feedback is lower than the preset instant confidence threshold, the target user's cross-session behavior tracking chain is triggered. The behavior tracking chain starts from the moment of this contact and has a preset decay window length as the validity period. It continuously captures all related behaviors generated by the user within the validity period. The related behaviors include delayed behaviors such as active search, deep browsing, adding to favorites and shopping cart, or cross-session conversion. When a related behavior is captured, the attribution weight is dynamically calculated based on the behavior type and the time interval between the current contact. The shorter the time interval and the higher the relevance between the behavior and the contacted content, the greater the attribution weight. Based on the attribution weights, calculate the weighted effect index of each strategy version within the decay window, and use the weighted effect index as input to update the attention model parameters and content generation model parameters; The strategy version identifier reached this time is bound and stored with the weighted effect metric for subsequent dynamic updates of the traffic bucket space.

7. The full-process digital promotion method based on intelligent agents according to claim 6, characterized in that, Before dynamically calculating the attribution weight based on the behavior type and the time interval between the current contact, the following steps are also included: Based on the aforementioned outreach impact record table, determine whether there are any historical outreach efforts that have been identified as generating positive conversions within the validity period; The dynamic calculation of attribution weights based on the time interval between the behavior type and the current contact includes the following steps: If it does not exist, the attribution weight is dynamically calculated based on the behavior type and the time interval between the current contact.

8. The full-process digital promotion method based on intelligent agents according to claim 7, characterized in that, After determining whether there are any historical contacts within the validity period that have been identified as generating a positive conversion, the process also includes the following steps: All historical touches identified as generating positive transitions are acquired. For each historical touch, the time interval between its touch time and the time of the associated behavior is calculated, and the intensity of the intent state transition corresponding to that touch is extracted. The transition intensity is determined by the ratio of the duration of the intent state after the touch to the duration of the intent state chaos before the touch. The time interval and the intensity of the intent state transition are input into a preset attribution weight calculation function to generate the contribution weight of this touch on the current associated behavior. The shorter the time interval and the greater the transition intensity, the higher the contribution weight. Normalize the contribution weights of all historical outreaches, and then distribute the effects of related behaviors to the corresponding strategy versions of each outreach according to the normalized weights. The allocated performance metrics are accumulated into the performance statistics of each strategy version for subsequent dynamic updates of traffic bucket space and optimization of attention model parameters.

9. A full-process digital promotion system based on intelligent agents, used to implement the full-process digital promotion method based on intelligent agents as described in any one of claims 1-8, characterized in that, include: A perceptual agent is configured to capture micro-behavioral events of a target user on a promotional session interface through a perception layer with real-time stream processing capabilities, timestamp each event, and combine them into a continuous event stream according to the time sequence. The intelligent agent is configured to call the temporal arbitration layer to perform slice analysis on the continuous event stream in a sliding window manner. If a conflict of intent tags pointing to different promotion targets is detected within a preset short window, it is determined to enter the intent chaos state and triggers a lightweight attention model to perform weight calculation on the micro-behavior sequence within the window, and outputs the current time-time target user's intent probability distribution for each candidate promotion target and its confidence score. A strategy generation agent is configured to trigger the decision layer to execute a tiered strategy generation based on a preset threshold range in which the confidence score falls: if it is higher than the first threshold, a precise promotion strategy targeting the highest probability target is generated; if it is between the first and second thresholds, a diversion and comparison promotion strategy containing multiple comparison versions is generated; if it is lower than the second threshold, a minimum guarantee strategy for pushing general content is generated, and a channel circuit breaker mechanism is triggered simultaneously to lock high-frequency reach channels. An execution agent is configured to call the execution layer to call the corresponding promotion channels in parallel via API according to the generated hierarchical strategy for content delivery, and continuously capture the feedback behavior of the target user after delivery, and feed the feedback behavior back to the perception layer as new micro-behavioral event input for online updating of the parameters of the attention model; as well as, Extraction agent, the extraction agent being configured for: Extract the trajectory of the target user's intent state before and after each touch. The trajectory of intent state evolution is generated in real time by the micro-behavioral events captured by the perception layer and determined by the temporal arbitration layer. It includes the duration of intent chaos within the preset window before the touch, whether the intent state changes after the touch, and the duration of the clear intent state after the change. The intent state evolution trajectory is used as the immediate impact feature of this outreach, and is stored together with the outreach time and policy version identifier in the outreach impact record table of the target user; When a related behavior is captured, the reach impact record table of the target user is invoked to identify all historical reach events that occurred within the validity period and resulted in a positive change in intent state after the reach. The positive change is used as candidate evidence that the reach has a long-term impact. The validity period is a preset attribution decay window length, and the positive change is the change in user intent state from chaotic to clear.

Images