A Marketing Strategy Iterative Optimization Method and System Based on User Profile Updates

Abstract

This invention relates to the field of marketing strategy optimization technology, and discloses a marketing strategy iterative optimization method and system based on user profile updates. This invention, through a designed conditional initialization mechanism, can obtain an initial profile feature vector and confidence weight vector with prior confidence assessment based on identity attributes. Through dynamic triggering based on confidence perception and dual-state collaborative updates, it prioritizes and accurately corrects features with low confidence and high uncertainty. Furthermore, this invention employs a confidence-based adaptive strategy generation mechanism, using confidence weights as a control signal input to the strategy generation model. This allows the model to dynamically adjust the randomness of output actions based on real-time perceived uncertainty. Thus, it proactively explores to collect information at low confidence levels and robustly utilizes it to gain benefits at high confidence levels. This addresses the problem of traditional methods failing to perceive and adapt to the confidence state of user profiles, leading to low decision-making efficiency.

Description

Technical Field

[0001] This invention relates to the field of marketing strategy optimization technology, and in particular to a marketing strategy iterative optimization method and system based on user profile updates. Background Technology

[0002] With the rapid development of internet and big data technologies, digital marketing has entered a stage of refined operation. Personalized marketing has become a core means to improve business conversion efficiency, enhance user stickiness, and increase user satisfaction, and is widely used in various scenarios such as e-commerce, services, and internet products. Its core concept lies in relying on big data collection and analysis technologies to comprehensively capture user behavior data across the entire user journey, integrating multi-dimensional information such as user historical behavior, interests, preferences, and basic attributes. Through feature extraction, cleaning, and fusion, a dynamically iterative user profile is constructed. Based on this user profile, the system can generate customized marketing strategies, thereby achieving a precise match between marketing actions and user needs.

[0003] Existing marketing decision-making systems based on user profiles suffer from inefficiencies in their strategy generation models. These models fail to perceive and adapt to the confidence levels of user profiles, resulting in low decision-making efficiency throughout the user lifecycle. Specifically, during the cold start phase with sparse data, the system struggles to generate high-confidence initial user profiles. At this stage, static / semi-static strategy models make blind decisions based on low-confidence profiles, leading to slow system startup. In the continuous iteration phase after data accumulation, user profiles are updated, and their confidence levels dynamically evolve. However, static / semi-static strategy models also lack the ability to respond to increases in confidence. Their update rate and decision rules are rigid, failing to smoothly transition from a high-exploration-intensity decision-making mode to a high-utilization-intensity decision-making mode based on the improvement in user profile confidence. This causes the feature space upon which model optimization relies to continuously deviate from the real-time evolving user state, resulting in overall low efficiency in marketing strategy iteration. Summary of the Invention

[0004] The main objective of this invention is to provide a marketing strategy iteration optimization method and system based on user profile updates, aiming to solve the technical problems in the prior art.

[0005] This invention proposes an iterative optimization method for marketing strategies based on user profile updates, comprising: Obtain the target user's identity attributes and current behavior observation data, and based on the condition initialization mechanism, obtain the target user's initial profile feature vector and initial confidence weight vector according to the identity attributes; Dynamic trigger analysis is performed based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector to determine whether to initiate a user profile update. If it is determined that a user profile update needs to be initiated, then the initial user profile feature vector and the initial confidence weight vector are jointly and incrementally optimized based on the current behavior observation data to obtain the updated profile feature vector and the updated confidence weight vector. Obtain the strategy generation model and the system context state vector, and generate a comprehensive state representation vector based on the updated profile feature vector and the system context state vector; The comprehensive state representation vector and the updated confidence weight vector are input into the strategy generation model. The strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector and outputs marketing action instructions. The marketing action instructions are executed, and the structured feedback data of the user on the instructions is collected through the tracking system and stored in the model training sample library; With the goal of learning a decision-making strategy with adaptive confidence, the strategy generation model is optimized and the optimization results are evaluated based on the model training sample library. The model optimization is then iteratively adjusted in a closed loop according to the evaluation results until the preset requirements are met and the corresponding personalized marketing strategy is output.

[0006] This application also provides a marketing strategy iteration and optimization system based on user profile updates, including: The feature acquisition module is used to acquire the identity attributes and current behavior observation data of the target user, and to acquire the initial profile feature vector and initial confidence weight vector of the target user based on the identity attributes according to the condition initialization mechanism. The trigger analysis module is used to perform dynamic trigger analysis based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector to determine whether to initiate user profile updates. The collaborative update module is used to perform collaborative incremental optimization on the initial user profile feature vector and the initial confidence weight vector based on the current behavior observation data if it is determined that a user profile update needs to be initiated, so as to obtain an updated profile feature vector and an updated confidence weight vector. The state fusion module is used to obtain the policy generation model and the system context state vector, and generate a comprehensive state representation vector based on the updated profile feature vector and the system context state vector. An adaptive decision-making module is used to input the comprehensive state representation vector and the updated confidence weight vector into the strategy generation model. The strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector and outputs marketing action instructions. The feedback collection module is used to execute the marketing action instructions and collect structured feedback data from users on the instructions through the tracking system and store it in the model training sample library. The iterative optimization module is used to optimize the strategy generation model based on the model training sample library with the goal of learning an adaptive confidence decision strategy, evaluate the optimization results, and perform closed-loop iterative adjustment of the model optimization according to the evaluation results until the preset requirements are met and the corresponding personalized marketing strategy is output.

[0007] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described marketing strategy iterative optimization method based on user profile updates.

[0008] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described marketing strategy iterative optimization method based on user profile updates.

[0009] The beneficial effects of this invention are as follows: By designing a conditional initialization mechanism, this invention can generate an initial profile feature vector and confidence weight vector with prior confidence assessment based on identity attributes and cross-domain knowledge transfer when the target user lacks historical data in this business domain. This solves the problem of failing to generate effective profiles in cold start and cross-domain scenarios and provides a quantitative basis for feature reliability for subsequent decisions. Through dynamic triggering and dual-state collaborative updates based on confidence perception, the system can prioritize and accurately correct features with low confidence and high uncertainty in an event-driven manner, while simultaneously improving the confidence level of effective features. This achieves optimal allocation of update timing and resources, as well as efficient and synchronous evolution of the user's cognitive state. Furthermore, this invention uses a confidence-based adaptive strategy generation mechanism, with confidence weights as a control signal input to the strategy generation model. This allows the model to dynamically adjust the randomness of output actions based on the uncertainty of real-time cognition, thereby actively exploring to collect information when confidence is low and robustly utilizing it to obtain benefits when confidence is high. This intelligently achieves a balance between high exploration intensity and high utilization intensity throughout the entire user lifecycle. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention.

[0012] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of this application.

[0013] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0015] like Figure 1 As shown, this application provides a marketing strategy iterative optimization method based on user profile updates, including: S1. Obtain the target user's identity attributes and current behavior observation data, and obtain the target user's initial profile feature vector and initial confidence weight vector based on the identity attributes according to the condition initialization mechanism; For example, this invention, through a design condition initialization mechanism, can determine whether a target user has historical records in the user profile database of this business domain based on the target user's identity attributes. If so, an initial profile feature vector and an initial confidence weight vector with prior confidence assessment are generated through cross-domain knowledge transfer or group profile prototype matching. If not, an initial profile feature vector is generated based on historical data, and an initial confidence weight vector is calculated based on the coverage and stability of historical data. This method allows the system to associate a quantified confidence weight with each user profile feature dimension in the initial stage. This not only solves the problem of traditional methods where new users or cross-business domain users lack historical data in this domain and cannot generate effective profiles, but also solves the problem that traditional methods, even if they obtain user profiles, lack a quantified assessment of the reliability of the user profile features themselves, resulting in the strategy production model being unable to distinguish the reliability of user profile information. S2. Based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector, perform dynamic trigger analysis to determine whether to initiate user profile update; For example, traditional user profile updates often employ fixed-period full updates, which suffers from severe lag and significant computational resource waste: ineffective updates occur when user interests remain unchanged, and timely responses are impossible when user interests change abruptly. To address this contradiction of inaccurate update timing and low resource efficiency, this invention transforms profile updates from time-driven to event-driven. This invention constructs a real-time dynamic trigger by extracting features from the user's initial profile feature vector, initial confidence weight vector, and current behavior observation data, and calculating a dynamic trigger value. This dynamic trigger value is compared to a preset baseline trigger threshold. If the dynamic trigger value is greater than the baseline threshold, the trigger condition is met, and the user profile update process is initiated; if the dynamic trigger value is not greater than the baseline threshold, the trigger condition is not met, and the update process is not initiated. This method achieves intelligent and resource-efficient update timing, overcoming the lag and resource waste of fixed-period updates while avoiding feature noise caused by blind updates. S3. If it is determined that a user profile update needs to be initiated, then based on the current behavior observation data, the initial user profile feature vector and the initial confidence weight vector are subjected to collaborative incremental optimization to obtain the updated profile feature vector and the updated confidence weight vector. For example, the present invention not only incrementally corrects the initial profile feature vector based on the current behavior observation data, but also synchronously updates the initial confidence weight vector. Specifically, the system increases the confidence weight of the updated features according to the information utility of the current behavior observation data. In this way, the user profile feature vector and the feature dimension confidence weight vector are synchronously iteratively updated as a set of associated state variables. This process simulates the convergence of the system's representation of user features (profile feature values) and its reliability assessment (confidence weight) from the initial rough estimate to the accurate estimate under the condition of continuously obtaining effective behavior observation data. S4. Obtain the strategy generation model and the system context state vector, and generate a comprehensive state representation vector based on the updated profile feature vector and the system context state vector; For example, a precise marketing decision must be based on a comprehensive reasoning process that considers both user profiles and the user's context. Relying solely on user profiles may lead to ineffective decisions at inappropriate times or when resources are insufficient. This invention generates a comprehensive state representation vector by concatenating updated user profile feature vectors with system context state vectors or by fusing them through a lightweight feature fusion network. This aims to build a complete and unified decision state view for the subsequent strategy generation model, enabling it to learn and model the interaction between user features and system context. The resulting comprehensive state representation vector not only encapsulates the latest user cognition and real-time contextual awareness but also serves as the sole and comprehensive input for the strategy generation model's reasoning and decision-making. It provides the necessary data foundation for the model to interpret the current situation, combine confidence information, and ultimately generate intelligent marketing instructions. S5. Input the comprehensive state representation vector and the updated confidence weight vector into the strategy generation model. The strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector and outputs marketing action instructions.

[0016] For example, this invention generates a deterministic baseline decision through a strategy generation model, and transforms the updated confidence weight vector into a noise intensity scale differentiated across dimensions. The noise intensity scale is then used to modulate standard random noise, generating a random perturbation vector to increase the randomness of the decision. This random perturbation vector is then injected into the baseline decision to form a final decision vector with adaptive randomness. Finally, the final decision vector is decoded into an executable marketing instruction. This modulation process ensures that when the user profile confidence is low, the strategy generation model outputs a higher entropy value for the action distribution to accelerate learning; conversely, when the confidence is high, the entropy value for the output action distribution is lower to stabilize revenue. This achieves a fully adaptive, intelligently balanced exploration and utilization marketing decision through the strategy generation model. S6. Execute the marketing action instruction, and collect the user's structured feedback data on the instruction through the tracking system and store it in the model training sample library; For example, this invention sends marketing action instructions to the execution engines of each user touchpoint, driving channels such as SMS push, APP pop-ups, and recommendation bar updates to achieve personalized outreach to target users; simultaneously, the end-to-end tracking system deployed on the user side is activated, which captures and records the user's full-dimensional feedback behavior to this marketing action with high granularity, forming standardized structured feedback data; subsequently, the structured feedback data is written to a distributed model training sample library for persistent storage in real time or near real time, providing a high-quality and timely feedback data source for subsequent model optimization; S7. With the goal of learning a decision-making strategy with adaptive confidence, the strategy generation model is optimized and the optimization results are evaluated based on the model training sample library. The model optimization is then iteratively adjusted in a closed loop according to the evaluation results until the preset requirements are met and the corresponding personalized marketing strategy is output. For example, unlike traditional optimization that aims to simply maximize returns, this invention explicitly aims to learn a decision-making strategy with adaptive confidence in the optimization of the strategy generation model. Based on the model training sample library, the optimization process drives the strategy generation model not only to learn actions that yield high returns, but also to learn how to adjust its exploration behavior based on the historical confidence information attached to the sample data. Specifically, when calculating the advantage function, an intrinsic reward term that is negatively correlated with the historical confidence weight is introduced. This term systematically improves the estimation of the advantage value of actions taken in historical states with low confidence. Thus, when updating parameters, the strategy model learns to tend to output a high-entropy (high-randomness) action distribution in low-confidence states. This invention addresses the risks associated with blind online optimization by introducing offline evaluation and a secure closed-loop control mechanism. It objectively evaluates the updated strategy generation model using an independent validation set that was not used in training, allowing for monitoring of the model's generalization ability and optimization performance. Closed-loop adjustments are made based on the evaluation results to ensure the correctness of the system's iteration direction, process stability, and result reliability. Ultimately, through iterative cycles of evaluation and adjustment, the system robustly converges to a stable state that continuously outputs efficient, secure, and personalized marketing strategies.

[0017] In one embodiment, step S1, which uses a conditional initialization mechanism to obtain the initial profile feature vector and initial confidence weight vector of the target user based on the identity attributes, includes: S11. Obtain the user profile database of this business domain, and query and match the user profile database based on the identity attributes of the target user to identify the user initialization mode identifier, wherein the user initialization mode identifier includes cross-domain transferable status, prototype matching required status and historical availability status of this domain. For example, this invention obtains a user profile database for the business domain. This user profile database is a structured database storing multi-dimensional feature vectors such as demographic information, historical behavioral characteristics, and preference tags of all registered users within the business domain. Based on the target user's identity attributes, such as unique identifiers like user ID and mobile phone number, the invention queries and matches the user profile database to obtain a user initialization pattern identifier. This user initialization pattern identifier represents the historical association status between the target user and the business domain. If a relevant record is found in the user profile database, it indicates that the target user is a returning user, and the user initialization pattern identifier is marked as historical for the domain. Available status; if no relevant record is found in the user profile database, it indicates that the target user is a new user in this business domain, and it is necessary to further determine whether they are an old user in other business domains (related domains). If a relevant record can be matched in the related domain database through identity attributes, the user initialization pattern identifier is marked as cross-domain transferable; if no relevant record is found in all known databases, it is determined to be a completely new user, and the user initialization pattern identifier is marked as prototype matching required. This method of classifying target users based on conditional judgment can ensure that when acquiring user profile features, they can be guided to the most suitable initialization path with the highest information utilization. S12. If the user initialization mode is identified as a cross-domain transferable state, then obtain the user association database and cross-domain feature mapping model of the associated business domain, and query the user association database based on the identity attribute to obtain the cross-domain user profile feature vector. The cross-domain user profile feature vector is input into the cross-domain feature mapping model to obtain the initial profile feature vector of this business domain, and the initial confidence weight vector is obtained based on the reconstruction error of the cross-domain feature mapping model. If the user initialization mode is identified as a prototype matching state, then based on the identity attribute, the prototype profile feature vectors and corresponding matching similarities of multiple group prototypes are matched from the user profile database, and the initial profile feature vector and initial confidence weight vector are obtained according to the prototype profile feature vector and the matching similarity. If the user initialization mode is identified as the historical availability status of this domain, then the historical profile feature vector is extracted from the user profile database as the initial profile feature vector, and the stability of each feature dimension in the initial profile feature vector is evaluated based on the user profile database to obtain the initial confidence weight vector.

[0018] For example, if the user initialization mode is identified as cross-domain transferable, this invention obtains a user association database and a cross-domain feature mapping model from related business domains. The user association database refers to a database used to store user profiles in other systems (such as e-commerce platforms, social applications, financial services, etc.) belonging to the same company as the marketing system involved in this invention but with different businesses. The cross-domain feature mapping model is a pre-trained machine learning model capable of mapping user profile features from one business domain to the feature space of another business domain, such as a neural network trained based on generative adversarial networks or canonical correlation analysis. The input of this model is the feature vector of the target user in business domain A, and the output is its estimated feature vector in business domain B (i.e., this business domain). This model is trained on a large amount of cross-domain user data and learns the semantic association and mapping between user features of different businesses. The model employs a method called "shooting pattern" to input cross-domain user profile feature vectors into a cross-domain feature mapping model. The model then uses forward propagation inference to output the initial profile feature vector of the target user in the current business domain. The initial profile feature vector refers to a feature dataset that carries user attribute information such as historical behavioral preferences, spending power, and interaction habits, including multiple user attribute feature values. In this process, the model calculates the error between the predicted value when outputting each feature dimension and a standard output of that dimension under ideal conditions obtained through methods such as autoencoders. The value of this error after being mapped by the sigmoid function is used as the initial confidence weight of that feature dimension. In this way, an intrinsic and interpretable reliability measure can be provided for user profiles transferred from external knowledge, enabling the system to know which transferred features are reliable and which need to be calibrated with subsequent behavioral data. If the user initialization mode is marked as requiring prototype matching, this invention uses observable information in the identity attributes, such as demographic attributes and registration channels, to match the K most similar group prototypes from the user profile database by calculating similarity. Each group prototype corresponds to a prototype profile feature vector and a prototype confidence weight vector. Based on the K prototype profile feature vectors and their corresponding matching similarities, a weighted average fusion is performed to generate an initial profile feature vector. This fusion result is then compensated for based on matching similarity: the fusion confidence of each feature dimension is multiplied by the attenuation coefficient of the highest matching similarity among all K prototypes for that dimension, thus obtaining the final initial confidence weight vector. If the user initialization mode is identified as the historical available state of this domain, this invention directly extracts the historical profile feature vector of the target user from the user profile database as the initial profile feature vector, and calculates the statistical variance of the historical value within a certain time window based on the historical value sequence of each feature dimension in the initial profile feature vector. The reciprocal of the variance is normalized and used as the confidence weight of that dimension, thereby obtaining the initial confidence weight vector. Through the aforementioned differentiated condition initialization mechanism, the system can intelligently select the optimal data source for three different types of target users: "cross-domain old users", "completely new users", and "local domain old users", and generate corresponding initial confidence weight vectors with clear statistical or model interpretations. This not only helps to solve the problem of lacking effective initial profiles when new users / cross-domain users are "cold-starting", but also provides a data foundation for subsequent confidence-aware user profile updates and adaptive strategy decisions.

[0019] In one embodiment, step S2, which involves dynamically triggering analysis based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector to determine whether to initiate a user profile update, includes: S21. Obtain the event type library and attribute-feature mapping table, and extract the target event and event occurrence timestamp from the current behavior observation data; For example, this invention provides standardized data mapping rules for subsequent event parameterization and feature parsing by acquiring an event type library and an attribute-feature mapping table. The event type library is a pre-established database that maps various user behaviors to discrete event types, such as browsing product details pages, adding items to the shopping cart, completing payment, and customer service inquiries, assigning basic importance weights and basic decay coefficients to each discrete event type. The attribute-feature mapping table is a pre-built and stored mapping data table that establishes the corresponding relationships between various attributes of the target event and predefined feature dimensions of the user profile. Furthermore, this invention employs a serial processing mechanism, extracting the earliest event from the current behavior observation data queue each time as the target event, and simultaneously extracting the event occurrence timestamp of the target event. S22. Based on the event type library, perform type matching and parameter query on the target event to obtain the target event code, basic importance weight and basic decay coefficient, and perform feature mapping on the target event based on the attribute-feature mapping table to obtain the real-time behavior feature vector; For example, the present invention obtains an event type vector by performing type matching and standardized encoding on the extracted target events based on an event type library: ,in, Represents an event type vector. Indicates the target event encoding. This indicates the basic importance weight corresponding to the target event. This represents the basic decay coefficient corresponding to the target event; and the target event is parsed to obtain the event attribute set. The parsed event attribute set is then mapped to a predefined feature dimension corresponding to the user profile through an attribute-feature mapping table to obtain a real-time behavior feature vector. S23. Dynamically compensate the basic attenuation coefficient according to the event occurrence timestamp to obtain the timeliness attenuation coefficient, and extract user attribute feature values ​​that are associated with the target event from the initial profile feature vector based on the target event code as associated feature components; For example, this invention obtains the last updated timestamp of the user profile based on the user's initial profile feature vector, and then... "Calculate the time-related decay coefficient, where, Indicates the time-related decay coefficient. Indicates the basic attenuation coefficient. Represents the decay rate constant. Indicates the timestamp of the event. This represents the timestamp of the profile update. This method enables adaptive sensitivity adjustment based on time intervals, meaning that when calculating dynamic trigger values ​​later, the trigger contribution of the current event can be dynamically adjusted according to the time interval of the user profile update. The longer the time since the last update, the trigger effectiveness of the same event will increase exponentially. This ensures that significant behaviors of users who have not been active for a long time can effectively trigger updates, while frequent similar behaviors of active users will not lead to redundant updates. This invention extracts user attribute feature values ​​that are related to the target event from the initial profile feature vector based on the currently encoded target event type. These extracted user attribute feature values ​​are then used as associated feature components to construct an associated profile feature vector. ,in, Represents the associated portrait feature vector. Indicates the first related to the target event Item-related feature components, Indicates the sequence number of the associated feature; S24. Obtain the preset range for each feature dimension, and calculate the original feature deviation based on the real-time behavior feature vector, the preset range, and the associated feature components; For example, the present invention obtains a preset range for each corresponding feature dimension in the real-time behavior feature vector and the associated portrait feature vector, and extracts the real-time feature value and associated feature component corresponding to each feature dimension from the real-time behavior feature vector and the associated portrait feature vector respectively, and then uses " "Calculate the original feature deviation, where, Indicates the first The original feature bias corresponding to each feature dimension Indicates the first Real-time feature values ​​corresponding to each feature dimension Indicates the first The associated feature components corresponding to each feature dimension Indicates the first Each feature dimension corresponds to an upper limit value within a preset range. Indicates the first Each feature dimension corresponds to a lower limit value within a preset range. The index represents the feature dimension. By calculating the original feature deviation, the deviation between the user's current real-time behavior and the user's long-term accumulated historical feature patterns can be quantified. This deviation is the core technical indicator for judging whether the user's interests, preferences, or state have undergone meaningful changes. S25. The original feature deviation is calibrated with confidence perception according to the initial confidence weight vector to obtain the calibrated feature deviation. The dynamic trigger value is calculated by using a weighted average algorithm based on the basic importance weight, the timeliness decay coefficient and the calibrated feature deviation corresponding to each feature dimension. For example, the present invention performs confidence-aware calibration on the original feature deviation degree one by one according to the confidence weight corresponding to each feature dimension in the initial confidence weight vector, and scales the original feature deviation degree by introducing a gain factor that is negatively correlated with the confidence weight of the feature dimension to obtain the calibrated feature deviation degree. This invention uses " "Calculate the dynamic trigger value, where, Indicates a dynamic trigger value. Indicates the number of feature dimensions. This indicates the basic importance weight corresponding to the target event. This represents the time-sensitivity decay coefficient corresponding to the target event. Indicates the first The calibration feature bias corresponding to each feature dimension This indicates the sequence number of the feature dimension. This method integrates three core factors: the inherent importance of the event, the sensitivity of the timeliness context, and the degree of user interest shift. By weighted averaging the deviation of each feature dimension, it quantifies the incremental information carried by the current event that is sufficient to drive the profile update. The larger the dynamic trigger value, the higher the necessity and urgency of triggering the update.

[0020] In one embodiment, step S3, which involves performing collaborative incremental optimization of the initial user profile feature vector and the initial confidence weight vector based on the current behavior observation data to obtain the updated profile feature vector and the updated confidence weight vector, includes: S31. Obtain a real-time behavior feature vector based on the current behavior observation data, and obtain a feature deviation vector based on the real-time behavior feature vector and the initial user profile feature vector; For example, the present invention extracts the corresponding real-time feature value and related feature component from the real-time behavior feature vector and the related profile feature vector for each related feature dimension, and calculates the difference between the real-time feature value and the related feature component for each related feature dimension to obtain the feature deviation vector. The feature deviation vector can be used to quantify the original difference magnitude between the current real-time behavior and the historical related features in each dimension. S32. Obtain the gain coefficient corresponding to the current event type, and perform a component-wise scalar multiplication operation on the feature deviation vector according to the gain coefficient to obtain the feature increment vector; For example, the present invention obtains a gain coefficient, which is a preset constant factor positively correlated with the current event type, and calculates the product of each feature deviation value in the feature deviation vector and the gain coefficient to obtain a feature increment vector. Since different event types carry different information credibility and value, the gain coefficient is introduced to calibrate and scale the original change amplitude represented by the difference between real-time and related feature values ​​according to the importance of the event, thereby transforming the feature changes reflected by the current real-time behavior from the original numerical difference into an effective information increment calibrated by the importance of the event. S33. Obtain the preset base learning rate, and obtain the adaptive learning rate based on the calibration feature deviation, the base learning rate, and the initial confidence weight vector; For example, the base learning rate obtained in this invention refers to the baseline parameter preset by the system to control the global speed of feature updates. The initial confidence weights corresponding to each feature dimension are extracted from the initial confidence weight vector, and the difference between "1" and the initial confidence weights is calculated to obtain the uncertainty factor. This value is used to measure the degree of uncertainty in the current understanding of that feature dimension. This invention obtains the adaptive learning rate by calculating the product of the base learning rate, the uncertainty factor, and the calibration feature deviation. This method allows adjustment of the base learning rate based on the uncertainty factor and calibration feature deviation of each feature dimension: a smaller uncertainty factor and more conventional behavior (i.e., a smaller calibration feature deviation) results in a lower adaptive learning rate, indicating a more conservative approach to user profile updates; a larger uncertainty factor and more abnormal behavior (i.e., a larger calibration feature deviation) results in a higher adaptive learning rate, indicating a more aggressive approach to user profile updates. S34. Based on the feature increment vector and the adaptive learning rate, the initial portrait feature vector is weighted and fused to update the portrait feature vector to obtain the updated portrait feature vector. For example, the present invention extracts the initial feature value corresponding to each feature dimension from the initial portrait feature vector, and then uses " "Calculate the feature update value, where, Indicates the first Feature update values ​​corresponding to each feature dimension Indicates the first The adaptive learning rate corresponds to each feature dimension. Indicates the first The initial feature values ​​corresponding to each feature dimension Indicates the first Feature increments corresponding to each feature dimension The index represents the feature dimension. This weighted moving average model treats the initial feature values ​​and the newly calculated feature increments as two information sources, linearly fusing them using an adaptive learning rate as the weight. When the adaptive learning rate is low, the calculation result mainly inherits the initial feature values, indicating a smoother and more conservative user profile update. When the adaptive learning rate is high, the calculation result tends to accept new increments, indicating a faster and more aggressive user profile update, and ensuring that the output value lies between the initial feature value and the target update value, thus facilitating the achievement of stable and smooth feature evolution. After performing the above incremental learning calculation on all related feature dimensions, all the obtained feature update values ​​are assembled in the original dimension order to form the updated profile feature vector. This vector is written back to the user profile database in real time, replacing the original initial profile feature vector, and used for subsequent strategy generation and the next trigger judgment. S35. Evaluate the information effectiveness of the current behavior observation data to obtain an information utility value, and update the initial confidence weight vector based on the information utility value, the feature increment vector and the adaptive learning rate to obtain an updated confidence weight vector. For example, this invention obtains the number of valid fields based on current behavior observation data, traverses a pre-defined list of data fields required for the current event type to obtain the total number of required fields, and calculates the ratio of the number of valid fields to the total number of required fields to obtain a completeness score. This invention also queries the event type library to obtain a basic credibility score corresponding to the current event type, which is pre-set based on the verifiability and business importance of the event itself. Simultaneously, it calculates the cosine similarity between the real-time behavior feature vector and the initial user profile feature vector to obtain a matching score, which reflects the degree of consistency between the current behavior and the user's recent profile. Then, it performs a weighted sum based on the completeness score, matching score, basic credibility score, and corresponding pre-defined weight coefficients to obtain an information utility value. This value serves as a global adjustment factor for subsequent confidence weight updates; the larger the value, the more reliable the current behavior data, and the greater the confidence improvement that can be driven. This invention uses " "Calculate the confidence weight increment, where, Indicates the first The confidence weight increment corresponding to each confidence weight. Indicates the information utility value. Indicates the first The adaptive learning rate corresponds to each feature dimension. Indicates the first Feature increments corresponding to each feature dimension This represents a preset, extremely small positive number, used to prevent the denominator from being zero. This indicates the preset confidence level update strength coefficient. This indicates the ordinal number of the feature dimension and confidence weight. The above formula defines the update rule for the confidence weight by coupling four technical variables: the information utility value serves as a pre-factor, globally permitting and scaling the update; its value originates from a direct assessment of data quality. The adaptive learning rate, as the second factor, couples feature updates with confidence updates, allowing the strength of feature updates to directly transmit and influence the evolution of confidence. The modulation term "..." "Taking the absolute value of the feature increment as input, its output value is inversely proportional to the absolute value of the feature increment. Its function is to transform the excessive feature correction amount into a signal that inhibits the increase of confidence. In this way, when the feature value changes drastically, it constrains the synchronous and rapid growth of confidence, and avoids premature saturation of cognitive certainty due to a single abnormal data. The confidence update strength coefficient is used to pre-set the overall time scale of confidence evolution, so as to achieve a stable, controlled incremental adjustment of confidence weight that is adapted to the feature update state." All the obtained confidence weight increments are superimposed with the initial confidence weights, and the superimposed results are assembled in the original dimensional order to form the updated confidence weight vector.

[0021] In one embodiment, step S5, which involves inputting the comprehensive state representation vector and the updated confidence weight vector into the strategy generation model, wherein the strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector, and outputs marketing action instructions, includes: S51. Input the comprehensive state representation vector into the policy generation model to obtain the baseline action vector, and obtain the noise scale vector based on the updated confidence weight vector. For example, the present invention inputs the comprehensive state representation vector into the backbone neural network of the strategy generation model. The network calculates through forward propagation and outputs a baseline action vector. The baseline action vector represents the original numerical representation of the optimal marketing action that the strategy generation model considers to be the best when it does not consider exploration and is based only on the current state information. Meanwhile, this invention obtains a noise scaling vector based on the updated confidence weight vector, which essentially establishes a positive correlation mapping between cognitive uncertainty and exploration intensity. Specifically, element-wise operations are performed on the updated confidence weight vector, and the calculation formula is as follows: ,in, Represents the first element in the noise scaling vector. The noise scale value corresponding to each feature dimension This represents the global noise intensity coefficient, used to control the baseline level of the overall exploration. This indicates the update of the confidence weight vector. The updated confidence weight value corresponds to each confidence weight. The index represents the feature dimension and confidence weight. The noise scale value corresponding to each feature dimension in the noise scale vector is the scaling factor of the exploration noise on the corresponding decision dimension. For feature dimensions with high confidence, the corresponding uncertainty factor is small, and the calculated noise scale value is also small. This means that the exploration noise added later on the decision component corresponding to this dimension should be weak, and the model decision should be more inclined to use existing certain knowledge. Conversely, for feature dimensions with low confidence weight, the noise scale value is large, which means that stronger random exploration should be carried out on this dimension to quickly obtain information and reduce uncertainty. S52. Obtain the standard Gaussian random noise vector, and multiply the standard Gaussian random noise vector with the noise scale vector element by element to obtain the modulated Gaussian noise vector. For example, this invention obtains a standard Gaussian random noise vector by sampling through a random number generator. The feature dimension of this vector is the same as that of the baseline action vector, and each component independently follows a standard Gaussian distribution with a mean of 0 and a standard deviation of 1. This vector is a purely random source that introduces decision-making randomness. This invention obtains a modulated Gaussian noise vector by performing element-wise multiplication of the standard Gaussian random noise vector with a noise scaling vector. This operation uses the noise scaling vector to differentially modulate the standard, isotropic random noise. After element-wise multiplication, the random values ​​of each feature dimension in the original standard Gaussian noise vector are scaled, so that the exploration noise in each decision dimension no longer has the same intensity, but is strictly linked to the system's confidence in the corresponding feature of that dimension: dimensions with high uncertainty have large noise variance and strong exploration disturbance; dimensions with low uncertainty have small noise variance and weak exploration disturbance. S53. Add the reference motion vector to the modulated Gaussian noise vector element by element to obtain the final motion vector; For example, this invention achieves the fusion of deterministic policy and controlled exploration by element-wise adding a baseline action vector to a modulated Gaussian noise vector to obtain the final action vector. The baseline action vector is the forward propagation calculation result of the policy generation model under a given integrated state representation vector. The modulated Gaussian noise vector is derived from standard Gaussian distribution random number sampling and is multiplied element-wise with a noise scaling vector. This multiplication ensures that each random component in the vector has its own independent conditional standard deviation, the size of which is determined by the corresponding dimension's noise scaling value. The addition of the two is equivalent to obtaining the baseline action vector. Each component is independently added with additive Gaussian noise whose standard deviation is the noise scale value corresponding to that component and whose mean is zero. From the perspective of the probabilistic model, this operation makes the final action vector follow a multidimensional Gaussian distribution with the reference action vector as the mean vector and the diagonal matrix composed of the squares of each component of the noise scale vector as the covariance matrix. The non-zero elements of the covariance matrix are directly determined by the noise scale vector, which is calculated by linear transformation of the updated confidence weight vector. Through the above operation, the dynamic and differentiated correlation between the statistical characteristics of the final output vector and the input signal is realized. S54. Perform a decoding mapping operation on the final action vector to convert it into a marketing action instruction; For example, the final action vector generated by the internal operation of the strategy generation model is usually a continuous, unconstrained numerical vector, while the marketing action instructions that can be executed by the actual system are often discrete or have a specific value range and format. Therefore, the present invention needs to use a decoder to complete the mapping from the numerical space to the action space, and convert the final action vector into a marketing action instruction by performing a decoding mapping operation on the final action vector.

[0022] In one embodiment, step S7, which aims to learn a decision-making strategy with adaptive confidence, optimizes the strategy generation model based on the model training sample library, evaluates the optimization results, and performs closed-loop iterative adjustments to the model optimization based on the evaluation results until preset requirements are met and a corresponding personalized marketing strategy is output, includes: S71. Extract the training dataset and independent validation set from the model training sample library, and update the parameters of the policy generation model based on the training dataset using the policy gradient algorithm to obtain the policy generation update model. For example, the present invention randomly divides the model training sample library into a training dataset and an independent validation set according to a preset ratio (e.g., 7:3). The training dataset is used to iteratively optimize the strategy to generate the model parameters, while the independent validation set does not participate in parameter updates and is specifically used to evaluate the model's generalization performance and confidence adaptive ability after each round of optimization. Each data sample contains historical behavior observation data, historical comprehensive state representation vector, historical marketing action instructions, historical confidence weight vector, and historical real-time revenue. Based on the training dataset, the system performs the following operations: A batch of data samples is randomly sampled from the training dataset. The historical comprehensive state representation vector and historical confidence weight vector of each sample in the batch are input into the current policy generation model. The model performs forward propagation and outputs the log probability of historical marketing action instructions. During this process, to guide the model to learn confidence-adaptive behavior, the calculation of the advantage function introduces an intrinsic reward term associated with the confidence state. Specifically, the advantage function estimator receives not only historical immediate returns, historical comprehensive state representation vectors, and historical marketing action instructions, but also historical confidence weight vectors. The calculated state-action value estimate is further supplemented with an average of the historical confidence weight vectors. The inversely proportional reward means that samples with lower confidence in historical data will have higher reward values ​​for their corresponding actions. This encourages the model to take actions that were previously highly rewarded in low-confidence states when facing them in the future during policy gradient updates. This drives the policy generation model to learn and establish a correlation between low-confidence states and high-value exploration actions, thereby enabling it to adaptively adjust its policy to more actively explore the action space with high-reward potential when facing uncertainty. Based on the calculated advantage function value and log probability, the gradient is calculated using the policy gradient theorem, and the parameters of the policy generation model are updated using an optimizer (such as Adam) to complete one round of optimization, resulting in the policy generation update model. S72. Based on the independent validation set and strategy, generate an updated model, perform model inference and profit calculation, and obtain the average profit index; For example, the present invention generates an updated model by inputting the historical integrated state representation vector and historical confidence weight vector of each sample in the independent validation set into the policy, obtains the action predicted by the model, obtains the predicted return through a lightweight return predictor, and calculates the arithmetic mean of the predicted returns of all validation samples as the average return index, which is used to evaluate the absolute effectiveness of the policy. S73. Extract the sample confidence weight vector from the independent validation set, and perform grouped statistical calculations on the sample confidence weight vector to obtain the confidence response difference index. For example, this invention extracts the sample confidence weight vector corresponding to each sample from the independent validation set and calculates its average value. Simultaneously, a preset confidence threshold is set, and samples with an average value higher than this threshold are classified into a high-confidence sample subset, while samples with an average value lower than this threshold are classified into a low-confidence sample subset. The following operations are performed on the two subsets respectively: the historical comprehensive state representation vector of each sample in the subset is input into the strategy to generate an updated model, obtaining the action probability distribution output by the model, and the entropy of this distribution is calculated. The higher the entropy value, the stronger the randomness of the model's decision in that state. Calculate the average value of the output distribution entropy of all samples in the high-confidence sample subset and the low-confidence sample subset respectively to obtain the average exploration intensity of high confidence and low confidence. Calculate the difference between the average exploration intensity of low confidence and the average exploration intensity of high confidence to obtain the confidence response difference index. Since an ideal strategy generation model with confidence adaptive capability should have a higher entropy of the output action probability distribution in the low-confidence state, the confidence response difference index should be positive. The larger the value, the stronger the ability of the strategy generation model to balance decision randomness adjustment according to the confidence state. S74. Obtain a preset first convergence threshold and a second convergence threshold, determine whether the average return index is greater than the first convergence threshold, and determine whether the confidence response difference index is greater than the second convergence threshold. If any judgment result is negative, then return to the step of updating the parameters of the policy generation model based on the training dataset using the policy gradient algorithm; If both judgments are yes, then based on the current behavior observation data and the system context state vector, the updated model is generated through the strategy to perform real-time reasoning and decision-making, and output a personalized marketing strategy. For example, if the judgment result does not reach the convergence target, new training data is used to continue the next round of parameter optimization for the strategy generation and update model. This loop will continue, driving the model to optimize returns while continuously strengthening its confidence-based adaptive decision-making ability. If the judgment result reaches the convergence target, the optimization iteration is terminated, and the current strategy generation and update model that meets the dual criteria is applied to online inference. Based on the current behavior observation data and the system context state vector, the system performs real-time inference decision-making through the strategy generation and update model to generate and output the final personalized marketing strategy.

[0023] This application also provides a marketing strategy iteration and optimization system based on user profile updates, including: The feature acquisition module is used to acquire the identity attributes and current behavior observation data of the target user, and to acquire the initial profile feature vector and initial confidence weight vector of the target user based on the identity attributes according to the condition initialization mechanism. The trigger analysis module is used to perform dynamic trigger analysis based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector to determine whether to initiate user profile updates. The collaborative update module is used to perform collaborative incremental optimization on the initial user profile feature vector and the initial confidence weight vector based on the current behavior observation data if it is determined that a user profile update needs to be initiated, so as to obtain an updated profile feature vector and an updated confidence weight vector. The state fusion module is used to obtain the policy generation model and the system context state vector, and generate a comprehensive state representation vector based on the updated profile feature vector and the system context state vector. An adaptive decision-making module is used to input the comprehensive state representation vector and the updated confidence weight vector into the strategy generation model. The strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector and outputs marketing action instructions. The feedback collection module is used to execute the marketing action instructions and collect structured feedback data from users on the instructions through the tracking system and store it in the model training sample library. The iterative optimization module is used to optimize the strategy generation model based on the model training sample library with the goal of learning an adaptive confidence decision strategy, evaluate the optimization results, and perform closed-loop iterative adjustment of the model optimization according to the evaluation results until the preset requirements are met and the corresponding personalized marketing strategy is output.

[0024] In one embodiment, the adaptive decision-making module includes: The preliminary decision-making unit is used to input the comprehensive state representation vector into the policy generation model to obtain the baseline action vector, and to obtain the noise scale vector based on the updated confidence weight vector. The noise modulation unit is used to acquire a standard Gaussian random noise vector and multiply the standard Gaussian random noise vector element-wise with the noise scaling vector to obtain a modulated Gaussian noise vector. The feature fusion unit is used to add the reference action vector to the modulated Gaussian noise vector element by element to obtain the final action vector; The instruction decoding unit is used to perform a decoding mapping operation on the final action vector and convert it into a marketing action instruction.

[0025] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described marketing strategy iterative optimization method based on user profile updates.

[0026] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described marketing strategy iterative optimization method based on user profile updates.

[0027] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0028] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0029] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A marketing strategy iterative optimization method based on user profile updates, characterized in that, include: Obtain the target user's identity attributes and current behavior observation data, and based on the condition initialization mechanism, obtain the target user's initial profile feature vector and initial confidence weight vector according to the identity attributes; Dynamic trigger analysis is performed based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector to determine whether to initiate a user profile update. If it is determined that a user profile update needs to be initiated, then the initial user profile feature vector and the initial confidence weight vector are jointly and incrementally optimized based on the current behavior observation data to obtain the updated profile feature vector and the updated confidence weight vector. Obtain the strategy generation model and the system context state vector, and generate a comprehensive state representation vector based on the updated profile feature vector and the system context state vector; The comprehensive state representation vector and the updated confidence weight vector are input into the strategy generation model. The strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector and outputs marketing action instructions. The marketing action instructions are executed, and the structured feedback data of the user on the instructions is collected through the tracking system and stored in the model training sample library; With the goal of learning a decision-making strategy with adaptive confidence, the strategy generation model is optimized and the optimization results are evaluated based on the model training sample library. The model optimization is then iteratively adjusted in a closed loop according to the evaluation results until the preset requirements are met and the corresponding personalized marketing strategy is output.

2. The marketing strategy iterative optimization method based on user profile updates according to claim 1, characterized in that, The step of obtaining the initial profile feature vector and initial confidence weight vector of the target user based on the identity attributes according to the conditional initialization mechanism includes: Obtain the user profile database of this business domain, and query and match the user profile database based on the identity attributes of the target user to identify the user initialization mode identifier, wherein the user initialization mode identifier includes cross-domain transferable status, prototype matching required status, and historical availability status of this domain. If the user initialization mode is identified as cross-domain transferable, then the user association database of the associated business domain and the cross-domain feature mapping model are obtained, and the user association database is queried based on the identity attribute to obtain the cross-domain user profile feature vector. The cross-domain user profile feature vector is input into the cross-domain feature mapping model to obtain the initial profile feature vector of this business domain, and the initial confidence weight vector is obtained based on the reconstruction error of the cross-domain feature mapping model. If the user initialization mode is identified as a prototype matching state, then based on the identity attribute, the prototype profile feature vectors and corresponding matching similarities of multiple group prototypes are matched from the user profile database, and the initial profile feature vector and initial confidence weight vector are obtained according to the prototype profile feature vector and the matching similarity. If the user initialization mode is identified as the historical availability status of this domain, then the historical profile feature vector is extracted from the user profile database as the initial profile feature vector, and the stability of each feature dimension in the initial profile feature vector is evaluated based on the user profile database to obtain the initial confidence weight vector.

3. The marketing strategy iterative optimization method based on user profile updates according to claim 1, characterized in that, The step of dynamically triggering analysis based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector to determine whether to initiate a user profile update includes: Obtain the event type library and attribute-feature mapping table, and extract the target event and event timestamp from the current behavior observation data; Based on the event type library, the target event is matched and parameters are queried to obtain the target event code, basic importance weight and basic decay coefficient. Based on the attribute-feature mapping table, the target event is feature-mapped to obtain a real-time behavior feature vector. The basic attenuation coefficient is dynamically compensated based on the event occurrence timestamp to obtain the timeliness attenuation coefficient, and user attribute feature values ​​associated with the target event are extracted from the initial profile feature vector based on the target event code as associated feature components. Obtain the preset range for each feature dimension, and calculate the original feature deviation based on the real-time behavior feature vector, the preset range, and the associated feature components; The original feature deviation is calibrated based on the initial confidence weight vector to obtain the calibrated feature deviation. A weighted average algorithm is then used to calculate the dynamic trigger value based on the basic importance weight, the timeliness decay coefficient, and the calibrated feature deviation corresponding to each feature dimension.

4. The marketing strategy iterative optimization method based on user profile updates according to claim 3, characterized in that, The step of performing collaborative incremental optimization on the initial user profile feature vector and the initial confidence weight vector based on the current behavior observation data to obtain the updated profile feature vector and the updated confidence weight vector includes: A real-time behavior feature vector is obtained based on the current behavior observation data, and a feature deviation vector is obtained based on the real-time behavior feature vector and the initial user profile feature vector; Obtain the gain coefficient corresponding to the current event type, and perform a component-wise scalar multiplication operation on the feature deviation vector based on the gain coefficient to obtain the feature increment vector; Obtain a preset base learning rate, and obtain an adaptive learning rate based on the calibration feature bias, the base learning rate, and the initial confidence weight vector; Based on the feature increment vector and the adaptive learning rate, the initial portrait feature vector is weighted and fused to update the portrait feature vector, resulting in an updated portrait feature vector. The information effectiveness of the current behavior observation data is evaluated to obtain an information utility value. Based on the information utility value, the feature increment vector, and the adaptive learning rate, the initial confidence weight vector is updated to obtain an updated confidence weight vector.

5. The marketing strategy iterative optimization method based on user profile updates according to claim 1, characterized in that, The step of inputting the comprehensive state representation vector and the updated confidence weight vector into the strategy generation model, wherein the strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector, and outputs marketing action instructions, includes: The integrated state representation vector is input into the policy generation model to obtain the baseline action vector, and the noise scale vector is obtained based on the updated confidence weight vector. Obtain a standard Gaussian random noise vector, and multiply the standard Gaussian random noise vector element-wise with the noise scaling vector to obtain a modulated Gaussian noise vector; The reference motion vector is added element by element to the modulated Gaussian noise vector to obtain the final motion vector; A decoding mapping operation is performed on the final action vector to convert it into a marketing action instruction.

6. The marketing strategy iterative optimization method based on user profile updates according to claim 1, characterized in that, The steps of optimizing the strategy generation model based on the model training sample library, evaluating the optimization results, and iteratively adjusting the model optimization based on the evaluation results until the preset requirements are met and the corresponding personalized marketing strategy is output, with the goal of learning an adaptive decision-making strategy based on confidence, include: The training dataset and independent validation set are extracted from the model training sample library, and the policy generation model is updated with parameters based on the training dataset using the policy gradient algorithm to obtain the policy generation update model. Based on the independent validation set and strategy, an updated model is generated for model inference and profit calculation to obtain the average profit index. The confidence weight vector of the samples is extracted from the independent validation set, and the confidence weight vector of the samples is grouped and statistically calculated to obtain the confidence response difference index. Obtain a preset first convergence threshold and a second convergence threshold, determine whether the average return index is greater than the first convergence threshold, and determine whether the confidence response difference index is greater than the second convergence threshold. If any judgment result is negative, then return to the step of updating the parameters of the policy generation model based on the training dataset using the policy gradient algorithm; If both judgments are yes, then based on the current behavior observation data and the system context state vector, the updated model is generated through the strategy to perform real-time reasoning and decision-making, and output a personalized marketing strategy.

7. A marketing strategy iterative optimization system based on user profile updates, characterized in that, include: The feature acquisition module is used to acquire the target user's identity attributes and current behavior observation data, and based on the condition initialization mechanism, acquire the target user's initial profile feature vector and initial confidence weight vector according to the identity attributes; The trigger analysis module is used to perform dynamic trigger analysis based on the current behavior observation data, the initial user profile feature vector, and the initial confidence weight vector to determine whether to initiate a user profile update. The collaborative update module is used to perform collaborative incremental optimization on the initial user profile feature vector and the initial confidence weight vector based on the current behavior observation data if it is determined that a user profile update needs to be initiated, so as to obtain an updated profile feature vector and an updated confidence weight vector. The state fusion module is used to obtain the policy generation model and the system context state vector, and generate a comprehensive state representation vector based on the updated profile feature vector and the system context state vector. An adaptive decision-making module is used to input the comprehensive state representation vector and the updated confidence weight vector into the strategy generation model. The strategy generation model dynamically modulates the randomness in the decision-making process based on the updated confidence weight vector and outputs marketing action instructions. The feedback collection module is used to execute the marketing action instructions and collect structured feedback data from users on the instructions through the tracking system and store it in the model training sample library. The iterative optimization module is used to optimize the strategy generation model based on the model training sample library with the goal of learning an adaptive confidence decision strategy, evaluate the optimization results, and perform closed-loop iterative adjustment of the model optimization according to the evaluation results until the preset requirements are met and the corresponding personalized marketing strategy is output.

8. The marketing strategy iterative optimization system based on user profile updates according to claim 7, characterized in that, The adaptive decision-making module includes: The preliminary decision-making unit is used to input the comprehensive state representation vector into the policy generation model to obtain the baseline action vector, and to obtain the noise scale vector based on the updated confidence weight vector. The noise modulation unit is used to acquire a standard Gaussian random noise vector and multiply the standard Gaussian random noise vector element-wise with the noise scaling vector to obtain a modulated Gaussian noise vector. The feature fusion unit is used to add the reference action vector to the modulated Gaussian noise vector element by element to obtain the final action vector; The instruction decoding unit is used to perform a decoding mapping operation on the final action vector and convert it into a marketing action instruction.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images