Private domain traffic activation system of full-chain data closed loop

Abstract

The invention provides a full-chain data closed-loop private domain traffic activation system, and relates to the technical field of data processing, and the system comprises an intention construction module which is used for obtaining user behavior data of multiple private domain user contacts, and obtaining a private domain traffic activation module based on the user behavior data; generating a dynamic track of intention evolution data which represents continuous evolution of the user intention and contains direction and rate; the activation decision module is used for generating a future continuation track through a prediction model according to the intention evolution data and determining an activation strategy according to the future continuation track; and the optimization module is used for executing the activation strategy and updating the prediction model based on user response data. According to the technical scheme, the problems of information loss and dynamic characteristic deficiency caused by static and discrete modeling in the prior art are solved, and the predictability, the accuracy and the self-adaptability of the activation decision are improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a private domain traffic activation system with a closed-loop data chain. Background Technology

[0002] With the deepening development of the digital economy, enterprises are increasingly focusing on the refined operation of private domain traffic to improve user activity, loyalty, and lifetime value. Currently, most related technical solutions adopt methods based on manual rules or traditional customer relationship management (CRM) systems, using static user tags and preset automated processes to activate and reach users in batches.

[0003] However, the aforementioned existing technologies have significant shortcomings in handling complex and ever-changing user behaviors. First, existing systems typically cannot integrate user behavior data from different private touchpoints, leading to "data silos" and loss of information dimensions, resulting in incomplete and one-sided user profiles. Second, these systems use discrete, static tags to model users, failing to capture the continuous and dynamic evolution of user intent; that is, they lack characterization of the "direction" and "rate" of intent changes, resulting in insufficient predictive ability. Their activation strategies and optimization processes are often "open-loop," lacking the ability to automatically and deeply optimize and adaptively adjust models and strategies based on real user feedback. Therefore, overcoming these shortcomings and providing a more accurate and dynamically adaptive activation system is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] To overcome the technical shortcomings of traditional data processing and decision-making systems, which suffer from information dimensionality loss and lack of dynamic characteristics due to the use of discrete and static modeling methods, when processing high-dimensional, continuous, and dynamically changing data such as user behavior, this application provides at least one private domain traffic activation system with a closed-loop data chain, including:

[0005] An intent building module is used to acquire user behavior data, wherein the user behavior data comes from at least two different private domain user touchpoints; based on the user behavior data, a dynamic trajectory representing the continuous evolution of user intent over time is generated; wherein the dynamic trajectory includes intent evolution data representing the direction and rate of user intent evolution at each state point;

[0006] The activation decision module is used to generate a continuation trajectory of the dynamic trajectory in the future time period through a prediction model based on the intent evolution data obtained from the latest state point of the dynamic trajectory; and to determine an activation strategy for intervening in the user intent evolution based on the difference between the continuation trajectory and the preset user target intent state.

[0007] An optimization module is used to execute the activation strategy and update the prediction model based on the user response data generated after executing the activation strategy.

[0008] As an optional implementation, generating a dynamic trajectory representing the continuous evolution of user intent over time includes:

[0009] Feature extraction is performed on the user behavior data to obtain user behavior features; the user behavior features are mapped to a high-dimensional user state manifold to determine the latest state point of the dynamic trajectory;

[0010] In response to receiving new user behavior data, the new user behavior data is parsed into an infinitesimal transformation acting on the latest state point, and the intention evolution data at the latest state point is generated.

[0011] Based on the intent evolution data, subsequent state points in the high-dimensional user state manifold are determined through state evolution processing, and the latest state point is connected with the subsequent state points to generate the dynamic trajectory.

[0012] As an optional implementation, generating the intent evolution data at the latest state point includes:

[0013] The newly added user behavior data is subjected to event type identification to determine the event type corresponding to the newly added user behavior data;

[0014] Based on the event type, the corresponding intent intensity value is retrieved from the weight library, and the intent intensity value is used to determine the modulus of the infinitesimal transformation;

[0015] Based on the event type, at least one intent evolution dimension corresponding to the high-dimensional user state manifold is determined;

[0016] The infinitesimal transformation is generated based on the intent intensity value and the at least one intent evolution dimension, and the intent evolution data is determined based on the infinitesimal transformation.

[0017] As an optional implementation, the optimization module is further configured to:

[0018] Based on the user response data, adjust the intent strength value in the weight library corresponding to the event type that triggered the user response data.

[0019] As an optional implementation, generating a continuation trajectory of the dynamic trajectory in future time periods using a prediction model based on the intent evolution data obtained from the latest state point of the dynamic trajectory includes:

[0020] Based on the latest state point of the dynamic trajectory, at least one local dynamic parameter associated with the latest state point is determined in the high-dimensional user state manifold, wherein the local dynamic parameter is used to characterize the local dynamic characteristics of the state manifold at the latest state point.

[0021] Based on the intent evolution data and the local dynamic parameters, an exponential mapping operation is performed on the latest state point to generate a continuation trajectory of the dynamic trajectory in the neighborhood of the latest state point.

[0022] As an optional implementation, the step of extracting features from the user behavior data to obtain user behavior features includes:

[0023] Structured attribute data is extracted from the user behavior data, and the structured attribute data is encoded into a static attribute vector;

[0024] Event sequence data is extracted from the user behavior data, and the event sequence data is processed by a pre-trained temporal coding model to generate a temporal feature vector;

[0025] Unstructured text data is extracted from the user behavior data, and the unstructured text data is processed by a pre-trained natural language processing model to generate semantic feature vectors;

[0026] The static attribute vector, the temporal feature vector, and the semantic feature vector are concatenated to form the user behavior feature.

[0027] As an optional implementation, mapping the user behavior features to a high-dimensional user state manifold and determining the latest state point of the dynamic trajectory includes:

[0028] The user behavior features are input into the encoder of the pre-trained variational autoencoder;

[0029] The encoder maps the user behavior features to latent space vectors in the high-dimensional user state manifold, and determines the latent space vectors as the latest state point of the dynamic trajectory.

[0030] As an optional implementation, querying and obtaining the corresponding intent strength value from the weight library includes:

[0031] Based on the event type, the corresponding basic intent strength value is obtained from the weight library;

[0032] Extract at least one data attribute from the newly added user behavior data to characterize user behavior features;

[0033] Based on the latest state point of the dynamic trajectory and the at least one data attribute, determine a dynamic adjustment coefficient for adjusting the basic intent intensity value;

[0034] The basic intent intensity value is adjusted using the dynamic adjustment coefficient to generate the intent intensity value.

[0035] As an optional implementation, determining the dynamic adjustment coefficient includes:

[0036] Based on the latest state point, extract features that characterize the local geometric topology of the state point in the high-dimensional user state manifold;

[0037] The at least one data attribute is input into an attribute encoding network to generate an attribute feature vector;

[0038] The features of the local geometric topology are fused with the attribute feature vector to generate a fused feature vector;

[0039] The fused feature vector is input into the coefficient generation model, which outputs the dynamically adjusted coefficients.

[0040] As an optional implementation, determining the activation strategy for intervening in the evolution of user intent includes:

[0041] Calculate the geometric deviation between the continuing trajectory and the user's target intention state, wherein the geometric deviation is used to quantitatively describe the spatial distance between the continuing trajectory and the user's target intention state in the high-dimensional user state manifold;

[0042] From a preset strategy library, multiple candidate activation strategies are selected based on the geometric deviation, where each candidate activation strategy corresponds to an intervention cost value;

[0043] For each candidate activation strategy, based on the geometric deviation and the intervention cost value, the corresponding comprehensive activation utility value is calculated through a preset utility function, wherein the utility function is used to comprehensively quantify the intervention effect and intervention cost of the activation strategy.

[0044] The candidate activation strategy with the highest overall activation utility value is determined as the activation strategy.

[0045] Compared with existing technologies that use discrete, static user tags for modeling and execute activation strategies through fixed rules, the technical solution provided in this application has at least the following advantages:

[0046] This application integrates user behavior data from multiple private domain touchpoints through an intent construction module and generates a continuous dynamic trajectory containing evolution direction and rate. Compared with the static label modeling of existing technologies, this modeling approach can make more comprehensive use of data and describe the dynamic changes in user intent, thereby reducing information loss caused by static data and models.

[0047] This application uses an activation decision module to predict a user's future trajectory based on the intent evolution data, and determines the activation strategy according to the difference between this trajectory and the target state. This ensures that the decision is based on the user's dynamic development trend, rather than just their historical state, thereby improving the accuracy and foresight of the activation strategy.

[0048] This application establishes a technical feedback loop by using user response data generated after executing the activation strategy to update the prediction model through an optimization module. This loop enables the system's prediction model to automatically adjust based on the actual activation effect, thus giving the system adaptive capabilities and allowing it to continuously optimize its decision-making basis to improve the effectiveness of subsequent activation strategies. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this application and, together with the specification, serve to explain the technical solutions of this application. It should be understood that the following drawings only show some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This illustration shows a schematic diagram of a private domain traffic activation system with a closed-loop data chain provided in an embodiment of this application;

[0051] Figure 2 The flowchart shown is a method for generating dynamic trajectories that represent the continuous evolution of user intent over time, according to an embodiment of this application.

[0052] Figure 3 A flowchart of a method for generating the intent evolution data at the latest state point, provided by an embodiment of this application, is shown.

[0053] Figure 4 A flowchart is shown, illustrating a method for generating a continuation trajectory of the dynamic trajectory in a future time period, according to an embodiment of this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0055] In this embodiment, some technical terms are defined as follows to ensure the clarity and consistency of the technical solution disclosed in this application:

[0056] Full-chain data closed loop: This refers to the system's ability to collect, store, process, and recirculate all types of user data throughout the entire lifecycle of private domain traffic, including all stages such as initial user contact, behavior tracking, tag improvement, conversion retention, and secondary activation, achieving closed-loop management of data flow. The full-chain data closed loop not only covers end-to-end management of the data link but also includes data anomaly feedback, effect review, and strategy self-optimization, ensuring data traceability and verifiability, and supporting continuous optimization of business decisions.

[0057] Private domain traffic refers to the user group that an enterprise or organization can independently reach, operate, manage, and repeatedly activate. Private domain traffic typically resides on platforms or channels controlled by the enterprise itself, such as WeChat official accounts, WeChat Work, APP membership systems, and mini-programs. The core of private domain traffic lies in user ownership and operational autonomy, distinguishing it from public domain traffic on external platforms.

[0058] Activation refers to using a series of precise online or offline marketing or outreach methods to prompt dormant or inactive private domain users to generate specific behavioral responses, such as clicks, inquiries, purchases, repeat purchases, and sharing. Activation methods can include push notifications, targeted offers, content outreach, and community interaction. Activation effectiveness data will be incorporated into the system's closed loop for continuous optimization.

[0059] See Figure 1 The diagram shown is a schematic of a private domain traffic activation system with a closed-loop data chain provided in an embodiment of this application. The system includes an intent building module 10, an intent building module 20, and an optimization module 30, wherein:

[0060] The intent construction module 10 is used to acquire user behavior data, wherein the user behavior data comes from at least two different private domain user touchpoints; based on the user behavior data, a dynamic trajectory representing the continuous evolution of user intent over time is generated; wherein the dynamic trajectory includes intent evolution data representing the direction and rate of user intent evolution at each state point;

[0061] The intent construction module 20 is used to generate a continuation trajectory of the dynamic trajectory in the future time period through a prediction model based on the intent evolution data obtained from the latest state point of the dynamic trajectory; and to determine an activation strategy for intervening in the evolution of user intent based on the difference between the continuation trajectory and the preset user target intent state.

[0062] The optimization module 30 is used to execute the activation strategy and update the prediction model based on the user response data generated after executing the activation strategy.

[0063] Module 10 is constructed in response to the above intent:

[0064] In practical implementation, the intent building module 10 can integrate multiple data interfaces for connecting and synchronizing data with different private domain user touchpoints. These private domain user touchpoints may include, but are not limited to: the backend server of the enterprise's own application, the cloud development platform of WeChat Mini Programs, the database of a customer relationship management system, and the API interfaces of social media platforms. Through these interfaces, the intent building module 10 can continuously acquire multi-source, heterogeneous user behavior data, such as structured transaction records and user profiles, as well as unstructured browsing logs and text comments, providing a foundation for achieving full-chain data analysis.

[0065] Furthermore, the core function of the intent construction module 10 is to generate a dynamic trajectory for each user, based on the acquired user behavior data, that represents the continuous evolution of their intent over time. This dynamic trajectory is a mathematical expression of user state that differs from traditional static labels; it aims to overcome the information loss caused by traditional discrete modeling methods. This dynamic trajectory treats the user's intent change process as a continuous motion in a high-dimensional space, thus reflecting the user's intent development process more realistically and completely.

[0066] To achieve quantitative analysis of the aforementioned dynamic trajectory, each state point of the trajectory contains a set of intent evolution data. This intent evolution data can be a set of parameters or a feature vector, used to quantitatively describe the instantaneous changing trend of the user's intent at that state point from a dynamic perspective. Specifically, it includes the direction of intent evolution, such as whether it tends towards purchase or churn, and the rate, i.e., the speed at which the intent changes. It should be noted that there can be various technical implementations for generating the dynamic trajectory and determining the intent evolution data, which will be described in detail in subsequent embodiments.

[0067] Module 20 is constructed in response to the above intent:

[0068] In practice, the intent construction module 20 receives dynamic trajectory and intent evolution data from the intent construction module 10 and makes decisions based on this data. This module contains a prediction model, which can employ various machine learning models familiar to those skilled in the art, such as time series prediction models, regression models, or deep learning networks.

[0069] The core function of the intent construction module 20 is to use the intent evolution data as key input and perform calculations through the prediction model to generate a continuation trajectory of the dynamic trajectory over a preset time period in the future. This continuation trajectory is a probabilistic prediction of the user's future intent direction. Subsequently, the module compares this continuation trajectory with a preset user target intent state, such as a state point or area defined in the state space as "complete purchase" or "high activity," to calculate the difference between the two.

[0070] Finally, the intent construction module 20 determines an activation strategy most suitable for intervening in the evolution of the current user intent based on the calculated differences. The activation strategy can be a set of multiple instructions, such as specifying activation channels like app push notifications or SMS messages, activation content like specific copy or coupon IDs, and activation timing like immediate execution or execution after 2 hours. There are various methods for determining the activation strategy, such as rule matching or optimization solutions; these will be described in more detail in subsequent embodiments.

[0071] Regarding the aforementioned optimization module 30:

[0072] In practical implementation, the optimization module 30 is a key execution and feedback component for realizing the closed-loop data flow of this system. On one hand, this module is configured to execute the activation strategy determined by the intent building module 20, which can call external communication gateways or service interfaces to achieve the final reach to the user.

[0073] On the other hand, and more importantly, this module is also used to capture and receive user response data through a data monitoring interface after the activation strategy is executed. This user response data objectively reflects the actual effect of the activation strategy, such as whether the user clicked on the push notification, claimed the coupon, or completed the purchase.

[0074] After acquiring the user response data, the optimization module 30 uses it as a feedback signal to update the prediction model in the intent construction module 20. This update process can be a model retraining or fine-tuning process, the purpose of which is to use the latest real data to correct the model parameters, making it more accurate in subsequent predictions. Through this "execution-feedback-update" process, the system forms a basic technical closed loop, giving it dynamic adaptive capabilities. Of course, the optimization mechanism of this application is not limited to this. In other embodiments, the optimization module 30 can also perform deeper optimizations on other parts of the system, which will be detailed later.

[0075] As an optional implementation, see [link to implementation details]. Figure 2 The flowchart illustrates a method for generating a dynamic trajectory representing the continuous evolution of user intent over time, according to an embodiment of this application, including steps S201 to S203, wherein:

[0076] S201: Extract features from the user behavior data to obtain user behavior features; map the user behavior features to a high-dimensional user state manifold to determine the latest state point of the dynamic trajectory;

[0077] S202: In response to receiving new user behavior data, the new user behavior data is parsed into an infinitesimal transformation acting on the latest state point, and the intention evolution data at the latest state point is generated.

[0078] S203: Based on the intent evolution data, determine the subsequent state points in the high-dimensional user state manifold through state evolution processing, and connect the latest state point with the subsequent state points to generate the dynamic trajectory.

[0079] In S201, the intent building module 10 first preprocesses and performs feature engineering on the raw user behavior data obtained from various private domain touchpoints. The feature extraction here aims to transform the multi-source, heterogeneous raw data into a unified, machine-understandable structured format, namely, user behavior features. For example, a user behavior feature can be a high-dimensional vector containing information from multiple dimensions, such as the user's static attributes, historical transaction statistics, and recent behavior frequency.

[0080] It should be noted that different feature extraction methods can be used for different types of data, such as structured data, event sequence data, and text data. More detailed explanations will be provided in subsequent embodiments of this application.

[0081] After obtaining user behavior features, the system maps them to a pre-constructed high-dimensional user state manifold. Here, the manifold is a mathematical concept, which can be understood as a continuous space capable of holding all possible user intent states. Through mapping, the user's complex features are transformed into a precise point in this space; this point is the latest state point of the dynamic trajectory, representing the user's most complete intent state at the current moment. There are various methods for mapping user features to the manifold, such as using a non-linear dimensionality reduction model based on deep learning.

[0082] When the system detects that a user has generated new behavior, i.e., receives newly added user behavior data, this application does not simply treat this new behavior as a new data point to update the user profile, but rather understands it as a "perturbation" or "push" exerted on the user's current intention state. Specifically, the system will "parse" this newly added user behavior data into a mathematical infinitesimal transformation.

[0083] In practice, the infinitesimal transformation here essentially quantifies the instantaneous impact of the new action on the user's intent. It comprises two core dimensions: direction and rate (or intensity). For example, the action of "adding a high-value item to the shopping cart" might be interpreted as a transformation pointing in the direction of "purchase" with high intensity; while "quickly scrolling through an unrelated product page" might be interpreted as a transformation with intensity close to zero. The final mathematical expression of this interpreted infinitesimal transformation is the intent evolution data mentioned above.

[0084] Of course, there are many ways to implement the specific logic and algorithms for parsing specific user behaviors into infinitesimal transformations.

[0085] After generating intent evolution data representing instantaneous trends, the intent construction module 10 calculates the next most likely position of the user's intent state through a state evolution process. This state evolution process can be understood as an integration or deduction operation performed on the curved space of the high-dimensional user state manifold. Starting from the latest state point and based on the direction and rate indicated by the intent evolution data, it calculates the subsequent state point that the user state should reach after a very small time unit.

[0086] Finally, the system mathematically connects the latest state point with the calculated subsequent state points, thus completing one generation or expansion of the dynamic trajectory. As user actions continue to occur, the system continuously repeats steps S202 and S203 to generate a smooth trajectory that can completely and continuously depict the entire evolution of the user's intent. Regarding the specific implementation of the state evolution processing, for example, an exponential mapping-based operation can be used; further details will be provided in subsequent embodiments of this application.

[0087] For example, in an e-commerce company's private domain operation system, the system is processing data related to user A.

[0088] Step 1: Determining the initial state point

[0089] At the initial time t0, the intent building module 10 first acquires and processes user A's historical user behavior data. This data is multi-source and heterogeneous, for example:

[0090] Data from the CRM system: User A, female, 28 years old, VIP level 3.

[0091] Data from the order system: Total spending in the past 90 days was 2,500 yuan, including purchases of items such as "dresses" and "high heels".

[0092] Data from the App's backend shows that there were 4 active days in the past 7 days, with a total of 35 products viewed, mainly concentrated in the "Handbags" and "Summer New Arrivals" categories.

[0093] The intent construction module 10 first performs feature extraction, processing the raw data into a structured user behavior feature vector. Then, through mapping—for example, inputting this feature vector into a pre-trained mapping model—it is mapped onto a high-dimensional user state manifold, thereby determining user A's latest state point at time t0, denoted as P0. This point P0 mathematically uniquely represents user A's current comprehensive intent state, which can be understood as: "a young female user with high value and high activity who is interested in fashion accessories."

[0094] Step 2: Generation of Intended Evolution Data

[0095] At the next moment t1, the system received a new user behavior data: "User A clicked on the 'New Summer Sandals' banner ad on the homepage of the App and browsed three different sandal products, with a total dwell time of 75 seconds."

[0096] At this point, the key role of the intent-building module 10 is not simply to record this browsing behavior, but to parse it into an infinitesimal transformation acting on state point P0. This parsing process will determine:

[0097] The "event type" of this behavior is "high intent browsing".

[0098] Its "intent intensity value" is relatively high.

[0099] Its "intention evolution dimension" mainly points to the two directions of "footwear category interest" and "recent purchase tendency" in the state manifold.

[0100] Based on the above analysis, the system ultimately generates intent evolution data at point P0, denoted as vector V0. This vector V0 mathematically quantitatively describes that, due to this browsing behavior, user A's intent state is evolving at a relatively fast rate towards "buying sandals".

[0101] Step 3: Generation and Extension of Dynamic Trajectories

[0102] After obtaining the latest state point P0 and intent evolution data V0, the intent construction module 10 performs state evolution processing. This processing can be a deductive operation based on V0 and P0, the purpose of which is to calculate the subsequent state point that user A should reach at time t1 on the high-dimensional user state manifold, denoted as P1.

[0103] Finally, the system mathematically connects points P0 and P1, for example, by using a smooth geodesic line, thus generating the first arc of the dynamic trajectory, from P0 to P1. At this point, the dynamic trajectory is generated and expanded. This new, latest state point P1 represents the evolution of user A's intention; its meaning has shifted from "interested in fashion accessories" to "having a high purchase intention for summer sandals." Subsequently, the system will make activation decisions based on this evolved new state P1.

[0104] In this way, the technical solution of this application records the user's behavior history completely and continuously as an analyzable and predictable dynamic trajectory, providing an unprecedented decision-making basis with dynamic characteristics for subsequent precise activation.

[0105] As an optional implementation, see [link to implementation details]. Figure 3 The flowchart illustrates a method for generating the intent evolution data at the latest state point according to an embodiment of this application, including S301 to S304, wherein:

[0106] S301: Perform event type identification on the newly added user behavior data to determine the event type corresponding to the newly added user behavior data;

[0107] S302: Based on the event type, query and obtain the corresponding intent intensity value from the weight library, the intent intensity value being used to determine the modulus of the infinitesimal transformation;

[0108] S303: Based on the event type, determine at least one intent evolution dimension corresponding to the high-dimensional user state manifold;

[0109] S304: Generate the infinitesimal transformation based on the intent intensity value and the at least one intent evolution dimension, and determine the intent evolution data based on the infinitesimal transformation.

[0110] Furthermore, this application establishes a structured and quantifiable processing flow to accurately transform a raw, qualitative user behavior into a mathematical object with a definite direction and modulus in a high-dimensional user state manifold, namely an infinitesimal variable, thereby overcoming the fuzziness and uncertainty of traditional analytical methods.

[0111] In its implementation, the intent building module 10 internally configures an event type definition library. This library predefines various standardized event types and includes rules for mapping raw user behavior data to these standard types. When new user behavior data, such as server logs or API callback data, enters the system, this step identifies it as a specific event type through rule matching or a classification model. For example, the identified event type could be "high-intent product browsing," "adding to cart," "claiming coupons," "posting positive reviews," or "initiating after-sales inquiries." This step achieves standardization and normalization of diverse user behaviors.

[0112] To quantify the influence of different behaviors on user intent, this embodiment's system also includes a weight library. This weight library stores the mapping relationship between event types and intent intensity values. The intent intensity value is a numerical value used to quantitatively describe the strength of the intent implied by an action; it is directly used to determine the magnitude of the infinitesimal transformation. For example, the weight library may include the following mapping relationship: the intensity value corresponding to "Add to Cart" is 5.0, and the intensity value corresponding to "Browse Products" is 1.0. This means that the driving force of the "Add to Cart" action on user intent is five times that of "Browse Products."

[0113] S303 above is used to determine the direction of the change in intent triggered by this user behavior. Based on the identified event type, the system determines that the behavior primarily affects one or more specific dimensions in the high-dimensional user state manifold. For example, in a fashion e-commerce scenario, the event type "browsing sandals" might be identified as simultaneously affecting three intent evolution dimensions: [seasonal preference], [footwear category preference], and [recent purchase intention]. By determining these dimensions, the system assigns a clear direction to this change in intent in the high-dimensional space.

[0114] After determining the "intensity" (magnitude) and "direction" (evolutionary dimension) of the intention change, this application mathematically synthesizes the two to generate the final infinitesimal transformation. In a preferred implementation, this infinitesimal transformation can be constructed as a vector. The magnitude of this vector is determined by the intention intensity value obtained in step S302, and its direction is formed by linearly combining one or more intention evolutionary dimensions determined in step S303 as basis vectors. The final generated mathematical object, in its data-driven representation, is the intention evolution data, which can be used for subsequent trajectory prediction and deduction.

[0115] In this way, a fuzzy "behavior parsing" process can be broken down into a clear, structured, and repeatable engineering process, ensuring that every user behavior can be transformed into a precise and computable intention evolution instruction, thereby greatly improving the accuracy and reliability of the entire dynamic trajectory modeling.

[0116] As an optional implementation, the optimization module 30 is further configured to:

[0117] Based on the user response data, adjust the intent strength value in the weight library corresponding to the event type that triggered the user response data.

[0118] In another preferred embodiment of this application, although the above-mentioned quantification of the intensity of user behavior intent is achieved through a weight library, if the correspondence in the weight library is static or preset, the system will not be able to dynamically and adaptively adjust its basic understanding of the importance of different user behaviors based on the actual operational effect, thereby limiting the system's long-term and deep adaptive capabilities.

[0119] Therefore, this embodiment enhances the functionality of the optimization module 30. In addition to updating the prediction model, the optimization module 30 also performs a deeper optimization task aimed at dynamically adjusting the weight library itself. Specifically, this process may include:

[0120] Step 1: Based on the user response data, adjust the intent intensity value in the weight library corresponding to the event type that triggered the user response data.

[0121] In practice, when the optimization module 30 receives a user response data, it first traces and associates it with the initial event type that triggered the response. For example, the system records that based on the action "event type = add to cart", it pushed an activation message to the user and eventually received "user response data = purchase completed".

[0122] Subsequently, the optimization module 30 will adjust the intent intensity value corresponding to the event type in the weight library based on the nature of the feedback result represented by the user response data, such as whether it is positive or negative feedback.

[0123] If the user response data is positive feedback, such as purchase, repeat purchase, or high-value interaction, the system will consider the initial "add to cart" event type to represent a stronger purchase intent than previously estimated. Therefore, optimization module 30 will appropriately increase the intent strength value corresponding to "add to cart" in the weight library, for example, from 5.0 to 5.1.

[0124] Conversely, if a series of activations triggered by "add to cart" ultimately lead to negative feedback, such as the user ignoring it or not taking any subsequent action for a long time, the system will determine that the actual intent strength of this event type may have been overestimated and will appropriately reduce its corresponding intent strength value.

[0125] Thus, this application constructs a dual-loop optimization mechanism. The first loop optimizes the decision-making model, enabling the system to learn how to make better decisions. The second loop, added in this embodiment, optimizes the weight library, allowing the system to learn and evolve its "basic cognitive rules" regarding user behavior. This mechanism transforms the system from a mere executor of fixed rules into one that continuously iterates and optimizes based on practical experience, fundamentally improving the system's long-term adaptability and activation effectiveness. Of course, there are various algorithms for adjusting the intent intensity value; for example, a dynamic adjustment coefficient can be calculated based on the user's state and behavioral attributes. Further details will be provided in subsequent embodiments.

[0126] As an optional implementation, see [link to implementation details]. Figure 4 The flowchart illustrates a method for generating a continuation trajectory of the dynamic trajectory in a future time period, according to an embodiment of this application, including steps S401 to S402, wherein:

[0127] S401: Based on the latest state point of the dynamic trajectory, determine at least one local dynamic parameter associated with the latest state point in the high-dimensional user state manifold, wherein the local dynamic parameter is used to characterize the local dynamic characteristics of the state manifold at the latest state point;

[0128] S402: Based on the intention evolution data and the local dynamic parameters, perform an exponential mapping operation on the latest state point to generate a continuation trajectory of the dynamic trajectory in the neighborhood of the latest state point.

[0129] In another embodiment of this application, conventional trajectory prediction methods typically assume that the evolution of a user's intent follows a globally uniform, linear pattern. However, the user's intent space is a complex, nonlinear, high-dimensional user state manifold, and its evolutionary pattern exhibits drastically different characteristics in different regions, i.e., when the user is in different intent states. If a simple linear extrapolation method is used for prediction, the predicted trajectory will deviate from the actual manifold surface, resulting in a large prediction error and failing to achieve accurate prediction.

[0130] Therefore, the technical problem that this embodiment aims to solve is: how to perform accurate, nonlinear trajectory prediction on a high-dimensional, curved state manifold, taking into account the local dynamic characteristics of the state points themselves.

[0131] In practice, the system assumes that the high-dimensional user state manifold is not uniform. The "terrain" of different regions, i.e., the inherent laws governing the evolution of user intent, are different. Therefore, before making predictions, the intent construction module 20 first analyzes the position of the user's current latest state point.

[0132] The system internally stores the local dynamic characteristics of different regions of the state manifold. These characteristics can be obtained through deep learning training on massive amounts of historical user trajectories. This step involves querying or calculating the local dynamic parameters describing the "dynamic characteristics" of the region where the current state point is located, based on the coordinates of the latest state point. These local dynamic parameters can be a set of numerical values, a matrix, or a tensor, used to quantitatively characterize the intention evolution trend of the region. For example, the acceleration tendency of intention changes in the region, the sensitivity coefficient to external intervention, or the coupling strength between different intention dimensions, etc.

[0133] After acquiring the intentional evolution data (which can be understood as the initial velocity vector) representing the current instantaneous change trend and the local dynamic parameters (which can be understood as terrain rules) representing the local spatial characteristics, this step will perform a nonlinear trajectory extrapolation.

[0134] In a preferred implementation, the derivation process is achieved through an exponential mapping operation. An exponential mapping operation is a concept in differential geometry that precisely maps a vector in the tangent space at a point—the intended evolution data—to a geodesic line on the manifold, i.e., the shortest path. In this application, this operation is modulated by the local dynamic parameters to ensure that the generated trajectory conforms to local dynamic laws.

[0135] The final result of this operation is a mathematically precise curve that always fits the surface of the high-dimensional user state manifold and extends in the neighborhood of the latest state point. This curve is the desired continuation trajectory.

[0136] For example, let's continue with user A. At time t1, their state is P1, and due to another click, a new intent evolution data V1 is generated. Now, the intent building module 20 needs to predict their future trajectory.

[0137] First, the module determines the local dynamic parameters of point P1 based on its location within the manifold. Analysis shows that point P1 is situated in a steep slope region where the intention changes very rapidly.

[0138] The module then takes point P1, vector V1, and this set of local dynamic parameters as input and performs an exponential mapping operation. Since the dynamic parameters indicate that this region is a steep slope, the operation will produce a continuation trajectory that is steeper and changes more rapidly than linear extrapolation.

[0139] This predictive trajectory accurately shows that, without external intervention, user A's purchase intention will surge rapidly in a short period, approaching the target state of "completing the purchase." This accurate prediction provides the strongest basis for decision-making regarding whether and how to intervene subsequently.

[0140] Thus, the technical solution of this application overcomes the technical defect of conventional linear prediction failing in complex nonlinear spaces. By introducing local dynamic characteristics and exponential mapping operations, it greatly improves the accuracy and reliability of predicting the user intention evolution trajectory.

[0141] As an optional implementation, the step of extracting features from the user behavior data to obtain user behavior features includes:

[0142] Structured attribute data is extracted from the user behavior data, and the structured attribute data is encoded into a static attribute vector;

[0143] Event sequence data is extracted from the user behavior data, and the event sequence data is processed by a pre-trained temporal coding model to generate a temporal feature vector;

[0144] Unstructured text data is extracted from the user behavior data, and the unstructured text data is processed by a pre-trained natural language processing model to generate semantic feature vectors;

[0145] The static attribute vector, the temporal feature vector, and the semantic feature vector are concatenated to form the user behavior feature.

[0146] In another embodiment, this application provides an innovative composite feature extraction scheme. Conventional feature extraction methods typically homogenize or flatten all types of user behavior data, for example, by simply performing numerical or one-hot encoding. This approach severely damages and loses the temporal correlation contained in event sequence data, as well as the deep semantic information contained in unstructured text data, resulting in insufficient accuracy in subsequent modeling.

[0147] Therefore, this embodiment aims to adopt the optimal encoding method adapted to different modalities of user behavior data in order to construct a composite user behavior feature that can retain the original information dimensions to the greatest extent.

[0148] In practice, the system first filters out the user's structured attribute data. This type of data is usually static or slowly changing, such as user registration information like age, gender, and region, membership level, and cumulative spending. For this type of data, the system can use conventional encoding methods, such as one-hot encoding for categorical features like gender and region, and normalization or standardization for numerical features like age and spending, ultimately generating a static attribute vector representing the user's basic attributes.

[0149] Next, the system processes the user's event sequence data. The core value of this type of data lies in its temporal order. For example, a user's action sequence in a single session might be "browsing product A -> browsing product B -> adding product A to the shopping cart". To capture this temporal relationship, this embodiment uses a pre-trained temporal coding model, such as a Recurrent Neural Network (RNN), a Long Short-Term Memory (LSTM) network, or a Transformer model, to process the event sequence. This model can compress the information of the entire sequence into a fixed-length temporal feature vector, which contains the dynamic patterns of user behavior and recent focus of interest.

[0150] Simultaneously, the system also processes users' unstructured text data, such as product reviews, social media posts, and chat logs with customer service. To uncover deeper semantics, this embodiment employs a pre-trained natural language processing model, such as BERT or a similar large language model, to process the text. This model can transform a piece of natural language text into a semantic feature vector that represents its emotional tone, key intentions, and themes.

[0151] Finally, the intent construction module 10 concatenates the static attribute vectors, temporal feature vectors, and semantic feature vectors—representing different aspects of user information—generated in the above three steps in a vector space. Through this concatenation, the system forms a higher-dimensional, more information-rich, and comprehensive description of the user's state—the final user behavior feature. This composite user behavior feature will serve as input for subsequent mapping to a higher-dimensional user state manifold.

[0152] In this way, this application overcomes the shortcomings of traditional flat feature extraction methods. By matching the most suitable AI model to the data of different modalities for encoding, valuable temporal and semantic information is preserved to the maximum extent, providing a high-quality data foundation for constructing accurate and reliable dynamic trajectories of user intent.

[0153] As an optional implementation, mapping the user behavior features to a high-dimensional user state manifold and determining the latest state point of the dynamic trajectory includes:

[0154] The user behavior features are input into the encoder of the pre-trained variational autoencoder;

[0155] The encoder maps the user behavior features to latent space vectors in the high-dimensional user state manifold, and determines the latent space vectors as the latest state point of the dynamic trajectory.

[0156] After obtaining complex user behavior features, the problem is how to map them into a low-dimensional space that can support subsequent continuous trajectory analysis. Conventional dimensionality reduction methods, such as Principal Component Analysis (PCA), are linear and cannot effectively handle the complex nonlinear relationships in user behavior features. While conventional autoencoders can perform nonlinear compression, the latent space they generate is often unstructured and discontinuous, potentially containing holes or tears, rendering trajectory interpolation or deduction in this space meaningless.

[0157] Therefore, this embodiment aims to map high-dimensional user behavior features into a structured, continuous, and smooth low-dimensional latent space to construct a high-dimensional user state manifold with excellent mathematical properties.

[0158] In practice, the system pre-trains a complete variational autoencoder (VAE) model using massive amounts of historical user data. A VAE model typically includes an encoder and a decoder. This step primarily utilizes the pre-trained encoder. After the intent building module 10 generates the user's composite user behavior features, these features are used as input and passed to the VAE's encoder. These features are typically a high-dimensional vector.

[0159] The encoder, typically a deep neural network, performs a series of nonlinear transformations and compressions on the input high-dimensional user behavior feature vector, ultimately mapping it to a low-dimensional latent space vector. This latent space vector represents the user's precise coordinates within the high-dimensional user state manifold.

[0160] For example, continuing with user A, this application generates a 1024-dimensional composite user behavior feature vector for user A. The system inputs this 1024-dimensional vector into the encoder of a pre-trained VAE. After calculation, the encoder outputs a 32-dimensional latent space vector. This 32-dimensional vector is determined as user A's latest state point P0 at time t0.

[0161] The significance of using VAE lies in its unique training mechanism, which aims to minimize reconstruction error while approximating the standard normal distribution of its latent space through regularization. This ensures that the generated high-dimensional user state manifold—the set of all possible latent space vectors—is continuous and smooth. This means that in this space, users with similar intentions will necessarily have similar spatial distances between their corresponding state points, and there are no "holes" that cannot be reached through smooth paths.

[0162] In this way, this application creatively solves the technical challenge of how to construct a mathematically "well-structured" user intent space. It ensures that all subsequent generation, deduction, and analysis of dynamic trajectories are carried out on a continuous, smooth, and meaningful mathematical foundation, providing a fundamental guarantee for the accuracy and reliability of the entire system.

[0163] As an optional implementation, querying and obtaining the corresponding intent strength value from the weight library includes:

[0164] Based on the event type, the corresponding basic intent strength value is obtained from the weight library;

[0165] Extract at least one data attribute from the newly added user behavior data to characterize user behavior features;

[0166] Based on the latest state point of the dynamic trajectory and the at least one data attribute, determine a dynamic adjustment coefficient for adjusting the basic intent intensity value;

[0167] The basic intent intensity value is adjusted using the dynamic adjustment coefficient to generate the intent intensity value.

[0168] The above text describes how a weight library quantifies the intent intensity of different event types. However, its approach of assigning a fixed intent intensity value to each event type fails to consider contextual information. For example, the same "add to cart" action, performed by a high-value active user and a price-sensitive dormant user, carries drastically different underlying intent intensity. Therefore, this embodiment aims to overcome the limitation of static weight libraries in reflecting contextual changes and achieve a dynamic and accurate calculation of intent intensity values ​​that are related to the user's current state and behavioral attributes.

[0169] This alternative implementation no longer uses the intensity value retrieved from the weight library directly, but instead uses it as a base value and calculates a dynamic adjustment coefficient based on the real-time context. After adjusting the base value, the final intent intensity value is obtained.

[0170] In practice, the system first retrieves a baseline intent strength value from the weight library based on the current event type. For example, the baseline intent strength value for the event type "add to cart" is 5.0.

[0171] At the same time, the system will extract the associated data attributes from the newly added user behavior data that triggered this calculation. These attributes describe the objective characteristics of the behavior. For example, for the behavior of "adding to cart", its data attributes may include: {product price: 599 yuan, product category: footwear, whether it is a promotional product: yes}.

[0172] The system will use information from two sources to calculate a dynamic adjustment coefficient:

[0173] The user's own state: that is, the latest state point of the dynamic trajectory, which represents whether the user is high-value or price-sensitive, active or dormant.

[0174] Objective attributes of behavior: These are the data attributes extracted in the previous step, which characterize the specific content of user behavior.

[0175] The system takes these two pieces of information as input, and outputs a dynamic adjustment coefficient through a preset calculation logic or model.

[0176] Finally, the calculated dynamic adjustment coefficients are used to adjust the obtained basic intent intensity value, for example, through multiplication or weighted operations, to obtain the final intent intensity value that reflects the real context.

[0177] Let's continue with user A as an example, and introduce a new user B for comparison. Assume the basic intent strength value for "add to cart" is 5.0.

[0178] For user A: her latest status point P1 indicates "high value, high activity"; she added a pair of summer sandals priced at 599 yuan (data attribute) to her shopping cart. Based on these inputs, the system calculates her dynamic adjustment coefficient to be approximately 1.2. Therefore, the final intent strength value of her action is 5.0 * 1.2 = 6.0.

[0179] For user B: his latest state point PB indicates that he is "price-sensitive and inactive"; he added the same item to his shopping cart. Based on these inputs, the system calculates that his dynamic adjustment coefficient is likely to be 0.7. Therefore, the final intention strength value of his behavior this time = 5.0 * 0.7 = 3.5.

[0180] In this way, the same "add to cart" action is precisely quantified into two completely different intent intensities. This embodiment overcomes the limitations of a static weight library by introducing a context-aware dynamic adjustment mechanism, making the system's quantification of user behavioral intent more accurate, intelligent, and personalized, thereby providing higher-quality data input for subsequent trajectory modeling and prediction.

[0181] As an optional implementation, determining the dynamic adjustment coefficient includes:

[0182] Based on the latest state point, extract features that characterize the local geometric topology of the state point in the high-dimensional user state manifold;

[0183] The at least one data attribute is input into an attribute encoding network to generate an attribute feature vector;

[0184] The features of the local geometric topology are fused with the attribute feature vector to generate a fused feature vector;

[0185] The fused feature vector is input into the coefficient generation model, which outputs the dynamically adjusted coefficients.

[0186] Conventional coefficient calculation models typically concatenate all input features, such as user state vectors and behavioral attributes, and then process them into a general model. The drawback of this approach is that it treats the latest state point, representing the user's state, merely as an ordinary vector, completely ignoring the rich geometric and topological information inherent in the high-dimensional user state manifold, resulting in significant information loss.

[0187] This embodiment provides a specific dual-channel fusion computing scheme. In its implementation, the system no longer simply uses the coordinates of the latest state point, but instead deeply analyzes its geometric properties in the manifold space. The system calculates and extracts a series of features that characterize the local geometric topology of the point. In different implementations, these features may include, but are not limited to:

[0188] Local curvature: Used to describe the degree of "bending" of the area where the point is located. A region with high curvature may mean that the user's intent is at a "turning point" that is changing rapidly.

[0189] Neighborhood density: Describes the density of other user state points in the vicinity of this point. A high-density area may represent a mainstream area of ​​user intent.

[0190] Geodesic distance: Used to calculate the shortest distance from a point along the manifold surface to a critical target state (such as the "purchase" state).

[0191] Meanwhile, the system processes the data attributes extracted from the newly added user behavior data in parallel. Since attributes such as price and category are usually conventional numerical or categorical data, the system can use an attribute encoding network, such as a standard multilayer perceptron (MLP) network, to encode them and transform them into a fixed-length attribute feature vector that can represent their inherent meaning.

[0192] After obtaining the features representing the local geometric topology ("geometric state information") and the attribute feature vectors representing the "event attribute information," this step effectively fuses these two types of features that come from different sources and have different properties. There are various ways to fuse them; for example, the two can be concatenated as vectors, or a more complex attention mechanism can be used for weighted fusion to generate a higher-level fused feature vector that simultaneously contains dual contextual information.

[0193] Finally, the system inputs the fused feature vector into a specialized coefficient generation model, such as one or more fully connected layers followed by an output unit. This model calculates and outputs a single scalar value, which is the desired dynamically adjusted coefficient that accurately reflects the current dual context.

[0194] For example, consider user A. To calculate the dynamic adjustment factor for their "add to cart" behavior:

[0195] Geometric Channel: The system analyzes its state point P1, extracts the features of its local geometric topology, and finds that the distance from this point to the geodesic of the purchase state is very close and the neighborhood density is high.

[0196] Attribute channel: The system inputs the data attributes of its behavior (price of 599 yuan, etc.) into the attribute encoding network to generate attribute feature vectors.

[0197] Fusion and Generation: The system fuses the two features mentioned above and inputs them into the coefficient generation model. The model understands that a user who is very close to making a purchase has a high intent to act on a moderately priced product, and therefore should be given a higher weight. Finally, the model outputs a dynamically adjusted coefficient of 1.2.

[0198] In this way, the aforementioned innovative dual-channel fusion framework overcomes the limitations of conventional models in processing and utilizing geometric information. By using the geometric topology of the user's state as one of the decision-making criteria, the system's understanding of the user context is greatly deepened, making the calculation of dynamic adjustment coefficients more accurate and reasonable, and ultimately significantly improving the accuracy of the entire intent modeling system.

[0199] As an optional implementation, determining the activation strategy for intervening in the evolution of user intent includes:

[0200] Calculate the geometric deviation between the continuing trajectory and the user's target intention state, wherein the geometric deviation is used to quantitatively describe the spatial distance between the continuing trajectory and the user's target intention state in the high-dimensional user state manifold;

[0201] From a preset strategy library, multiple candidate activation strategies are selected based on the geometric deviation, where each candidate activation strategy corresponds to an intervention cost value;

[0202] For each candidate activation strategy, based on the geometric deviation and the intervention cost value, the corresponding comprehensive activation utility value is calculated through a preset utility function, wherein the utility function is used to comprehensively quantify the intervention effect and intervention cost of the activation strategy.

[0203] The candidate activation strategy with the highest overall activation utility value is determined as the activation strategy.

[0204] After predicting the user's continued trajectory, how can we intelligently select an optimal activation strategy? Conventional methods are usually based on a simple rule, such as executing a preset action if the predicted purchase probability is below a certain threshold. This approach is coarse-grained, failing to finely quantify the deviation between the predicted trajectory and the target, and also failing to comprehensively consider the costs of different intervention strategies themselves.

[0205] This embodiment provides an activation strategy determination scheme based on utility optimization. The strategy selection problem is transformed into an optimization problem with cost constraints.

[0206] In practical implementation, the intent construction module 20 first needs to quantitatively describe the gap between the predicted continuation trajectory and the final user target intent state. Since the user's state exists on a high-dimensional, curved manifold, this gap is not a simple straight-line distance. The geometric deviation is a measure that more accurately reflects the spatial distance between the two within the high-dimensional user state manifold. In a preferred implementation, this geometric deviation can be the geodesic distance from the endpoint of the continuation trajectory to the region where the user target intent state is located.

[0207] The system has a pre-configured strategy library containing various executable activation strategies, such as {Strategy A: Send a 10 yuan coupon, Strategy B: Push a new product introduction article, Strategy C: Send an SMS reminder for unchecked shopping cart, Strategy D: Temporarily do not disturb}. Importantly, each strategy corresponds to an intervention cost value, which can be multi-dimensional. For example, it includes both financial costs such as the coupon amount and SMS fees, and the "disturbing cost" to the user; for instance, the disturbing cost of pushing a message is higher than not disturbing the user.

[0208] The system can select suitable candidate activation strategies from the strategy library based on the geometric deviation calculated in the steps. For example, when the deviation is large, strategies with stronger intervention are selected; when the deviation is small, strategies with weaker intervention or lower cost are selected.

[0209] For each candidate activation strategy, the system evaluates its "cost-effectiveness" using a utility function. The purpose of this utility function is to comprehensively quantify the intervention effect and cost of an activation strategy. Its inputs include at least:

[0210] Predicting the intervention effect: This is related to the geometric deviation. The greater the degree to which a strategy can bring the user's trajectory closer to the target state, the better its effect.

[0211] Intervention cost: This refers to the intervention cost associated with the strategy.

[0212] This utility function (e.g., utility = expected benefit - expected cost) calculates a final, single scalar value for each candidate activation strategy, namely the overall activation utility value.

[0213] Finally, the intent building module 20 selects the one with the highest overall activation utility value among all candidate activation strategies and uses it as the final output and execution activation strategy.

[0214] For example, consider user A. The system predicts that the geometric deviation of their continued trajectory from the "purchase" target state is very small. Based on this, the system selects two low-cost candidate strategies from the strategy library: {Strategy B: Push new product introduction articles, Strategy D: Do not disturb for now}.

[0215] For strategy B, the utility function calculates its overall activation utility value to be 1.5.

[0216] For strategy D, the utility function calculates its overall activation utility value to be 0.5, because not disturbing others also has its value, thus avoiding the risk of user churn.

[0217] Ultimately, the system selected strategy B, which has a higher utility value, for user A.

[0218] This elevates the process of determining the activation strategy from a fuzzy, rule-based matching to a refined, cost-benefit-based mathematical optimization. This ensures that every intervention taken by the system is the most cost-effective choice under the current circumstances, thereby maximizing the overall return on investment in activation.

[0219] For example, the utility function can be structurally designed as a multi-module, step-by-step computational process, which aims to deconstruct the abstract concept of "utility" into a series of quantifiable and computable intermediate steps. This process may include an expected benefit assessment component, an intervention cost assessment component, and a utility value synthesis component.

[0220] For the expected return assessment component: This component's function is to evaluate the potential positive business benefits of a given candidate activation strategy. Its internal processing logic may include:

[0221] To predict the intervention effect of the strategy, this component first calls an internal "strategy effect prediction model." It takes the user's latest state point, the current geometric deviation, and the candidate activation strategy itself as input. The model outputs a predicted "post-intervention deviation." This output value represents the system's prediction of a new, reduced deviation between the user's trajectory and the target state after implementing the strategy.

[0222] The component calculates the base return by subtracting the predicted "post-intervention bias" from the initial geometric bias, resulting in a "bias reduction." This "bias reduction" is considered the base return brought by the strategy; the larger the value, the better the strategy's performance.

[0223] User value weighting is applied. To reflect the logic that "the same effect is achieved with more valuable users, resulting in higher returns," this component queries and obtains a "user value coefficient" based on the user's latest status. For example, this coefficient can be calculated based on factors such as the user's historical total spending and activity level. Then, this "user value coefficient" is multiplied by the "deviation reduction" calculated in the second step to obtain the final, value-weighted expected return.

[0224] For the intervention cost assessment component: This component's function is to evaluate the overall cost required for the same candidate activation strategy. Its internal processing logic may include:

[0225] To obtain the financial cost, this component directly queries the policy library for the financial cost portion of the intervention cost value corresponding to the candidate activation policy. For example, for the policy "send a 10 yuan coupon", its financial cost is 10.

[0226] To calculate the cost of user disruption and achieve sustainable operation, this component innovatively introduces the concept of "disruption cost." The system quantifies this cost based on the "intrusiveness" of the strategy. For example, the disruption cost of an "in-app pop-up reminder" might be set at 2; the disruption cost of an "SMS push notification" might be set at 5; while the disruption cost of the "no disruption" strategy is 0.

[0227] The component synthesizes a comprehensive intervention cost by weighting and summing the "financial cost" obtained in the first step with the "user disturbance cost" calculated in the second step. This cost comprehensively reflects the price incurred in implementing the strategy.

[0228] For the utility value synthesis component, this component performs the final calculation. It receives the expected benefits output by the "Expected Benefit Assessment Component" and the total intervention cost output by the "Intervention Cost Assessment Component".

[0229] The calculation logic is as follows: subtract the comprehensive intervention cost from the expected benefits, and the final calculation result is the comprehensive activation utility value of the candidate activation strategy.

[0230] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A private domain traffic activation system with a closed-loop data chain, characterized in that, include: An intent building module is used to acquire user behavior data, wherein the user behavior data comes from at least two different private domain user touchpoints; based on the user behavior data, a dynamic trajectory representing the continuous evolution of user intent over time is generated; wherein the dynamic trajectory includes intent evolution data representing the direction and rate of user intent evolution at each state point; The activation decision module is used to generate a continuation trajectory of the dynamic trajectory in the future time period through a prediction model based on the intent evolution data obtained from the latest state point of the dynamic trajectory; and to determine an activation strategy for intervening in the user intent evolution based on the difference between the continuation trajectory and the preset user target intent state. An optimization module is used to execute the activation strategy and update the prediction model based on the user response data generated after executing the activation strategy.

2. The private domain traffic activation system with a closed-loop data chain according to claim 1, characterized in that, The generated dynamic trajectory representing the continuous evolution of user intent over time includes: Feature extraction is performed on the user behavior data to obtain user behavior features; the user behavior features are mapped to a high-dimensional user state manifold to determine the latest state point of the dynamic trajectory; In response to receiving new user behavior data, the new user behavior data is parsed into an infinitesimal transformation acting on the latest state point, and the intention evolution data at the latest state point is generated. Based on the intent evolution data, subsequent state points in the high-dimensional user state manifold are determined through state evolution processing, and the latest state point is connected with the subsequent state points to generate the dynamic trajectory.

3. The private domain traffic activation system with a closed-loop data chain according to claim 2, characterized in that, The intention evolution data at the latest state point includes: The newly added user behavior data is subjected to event type identification to determine the event type corresponding to the newly added user behavior data; Based on the event type, the corresponding intent intensity value is retrieved from the weight library, and the intent intensity value is used to determine the modulus of the infinitesimal transformation; Based on the event type, at least one intent evolution dimension corresponding to the high-dimensional user state manifold is determined; The infinitesimal transformation is generated based on the intent intensity value and the at least one intent evolution dimension, and the intent evolution data is determined based on the infinitesimal transformation.

4. The private domain traffic activation system with a closed-loop data chain according to claim 3, characterized in that, The optimization module is also used for: Based on the user response data, adjust the intent strength value in the weight library corresponding to the event type that triggered the user response data.

5. A private domain traffic activation system with a closed-loop data chain according to claim 2, characterized in that, The step of generating a continuation trajectory of the dynamic trajectory in future time periods using a prediction model based on the intention evolution data obtained from the latest state point of the dynamic trajectory includes: Based on the latest state point of the dynamic trajectory, at least one local dynamic parameter associated with the latest state point is determined in the high-dimensional user state manifold, wherein the local dynamic parameter is used to characterize the local dynamic characteristics of the state manifold at the latest state point. Based on the intent evolution data and the local dynamic parameters, an exponential mapping operation is performed on the latest state point to generate a continuation trajectory of the dynamic trajectory in the neighborhood of the latest state point.

6. The private domain traffic activation system with a closed-loop data chain according to claim 2, characterized in that, The method for extracting features from the user behavior data to obtain user behavior features includes: Structured attribute data is extracted from the user behavior data, and the structured attribute data is encoded into a static attribute vector; Event sequence data is extracted from the user behavior data, and the event sequence data is processed by a pre-trained temporal coding model to generate a temporal feature vector; Unstructured text data is extracted from the user behavior data, and the unstructured text data is processed by a pre-trained natural language processing model to generate semantic feature vectors; The static attribute vector, the temporal feature vector, and the semantic feature vector are concatenated to form the user behavior feature.

7. A private domain traffic activation system with a closed-loop data chain according to claim 6, characterized in that, The step of mapping the user behavior features to a high-dimensional user state manifold and determining the latest state point of the dynamic trajectory includes: The user behavior features are input into the encoder of the pre-trained variational autoencoder; The encoder maps the user behavior features to latent space vectors in the high-dimensional user state manifold, and determines the latent space vectors as the latest state point of the dynamic trajectory.

8. The private domain traffic activation system with a closed-loop data chain according to claim 3, characterized in that, The step of querying and obtaining the corresponding intent strength value from the weight library includes: Based on the event type, the corresponding basic intent strength value is obtained from the weight library; Extract at least one data attribute from the newly added user behavior data to characterize user behavior features; Based on the latest state point of the dynamic trajectory and the at least one data attribute, determine a dynamic adjustment coefficient for adjusting the basic intent intensity value; The basic intent intensity value is adjusted using the dynamic adjustment coefficient to generate the intent intensity value.

9. A private domain traffic activation system with a closed-loop data chain according to claim 8, characterized in that, Determining the dynamic adjustment coefficient includes: Based on the latest state point, extract features that characterize the local geometric topology of the state point in the high-dimensional user state manifold; The at least one data attribute is input into an attribute encoding network to generate an attribute feature vector; The features of the local geometric topology are fused with the attribute feature vector to generate a fused feature vector; The fused feature vector is input into the coefficient generation model, which outputs the dynamically adjusted coefficients.

10. A private domain traffic activation system with a closed-loop data chain according to claim 5, characterized in that, The activation strategy determined for intervening in the evolution of user intent includes: Calculate the geometric deviation between the continuing trajectory and the user's target intention state, wherein the geometric deviation is used to quantitatively describe the spatial distance between the continuing trajectory and the user's target intention state in the high-dimensional user state manifold; From a preset strategy library, multiple candidate activation strategies are selected based on the geometric deviation, where each candidate activation strategy corresponds to an intervention cost value; For each candidate activation strategy, based on the geometric deviation and the intervention cost value, the corresponding comprehensive activation utility value is calculated through a preset utility function, wherein the utility function is used to comprehensively quantify the intervention effect and intervention cost of the activation strategy. The candidate activation strategy with the highest overall activation utility value is determined as the activation strategy.

Images