An AI scenario perception-based dynamic personalized interaction system and method

Abstract

The application relates to the field of artificial intelligence and man-machine interaction technology and discloses a dynamic individualized interaction system and method based on AI scene sensing, which comprises the following steps: a multi-modal scene sensing module collects behavior, asset and environment data and maps the data to a unified feature space; an intelligent decision and interaction engine module switches detection and adaptation modes based on confidence, uses multi-agent game allocation resources and a power of purchase visual mapping to generate a strategy; a dynamic rendering execution module uses an affine transformation and a shader parameter to render an interface and feed back a structured signal based on a server-side driving architecture; and a feedback optimization closed loop module calculates a composite reward, updates strategy parameters through online learning and calibrates a model. The application balances business conflicts through multi-agent game, adaptively detects intentions and a power of purchase visual mapping through confidence difference, and realizes the collaborative optimization of dynamic layout and intuitive interaction.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and human-computer interaction technology, specifically to a dynamic personalized interaction system and method based on AI context awareness. Background Technology

[0002] With the rapid development of mobile internet technology, points malls and various e-commerce platforms have become core channels for enterprises to conduct user operations and asset monetization. Given the extremely limited physical constraints of mobile terminal screen size, efficiently utilizing every pixel's display area to present the most valuable goods or services has become crucial for improving platform conversion rates. This not only involves accurately capturing individual user preferences but also finding a dynamic balance among complex business objectives such as inventory turnover, cash flow, and user retention. Therefore, building an intelligent screen layout and interaction decision-making mechanism is particularly important.

[0003] Existing screen layout generation technologies mainly rely on pre-defined static template rules or traditional recommendation systems based on collaborative filtering and deep learning. The typical process is as follows: the system first collects users' historical click and browsing behavior data, and uses a recommendation algorithm model to predict the click-through rate of candidate products; then, the system sorts the products according to their predicted probabilities and fills the top-ranked products into fixed UI template slots; some systems may introduce manually set weighting rules to apply simple traffic bias to products of specific business types, and finally, the generated list is presented to the user in a standardized graphic and text format through the client-side rendering engine.

[0004] While existing technologies can achieve personalized content recommendations to some extent, they still have significant shortcomings in handling multi-objective dynamic game scenarios and complex human-computer interaction scenarios. First, traditional recommendation algorithms are usually optimized based on a single objective function (such as maximizing click-through rate), lacking a dynamic coordination mechanism for multiple and often conflicting business objectives such as inventory depth and asset consumption rate. This makes it difficult for the system to satisfy user interests while taking into account the overall business benefits of the enterprise. Second, existing technologies often lack proactive intent detection mechanisms in the stage of ambiguous user intent or cold start, and can only forcibly output low-confidence recommendation results based on sparse data. This mode is prone to noise interference and degrades the user experience. In addition, at the interface presentation layer, most existing technologies adopt linear visual rendering logic. The price of goods and the user's asset capacity are only compared through numerical text, failing to establish an intuitive psychological mapping between numerical purchasing power and visual salience. This requires users to perform additional mental calculations to judge the ease of acquisition while browsing, and cannot effectively use visual stimuli to lower the decision-making threshold or stimulate consumption impulses in a critical state. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a dynamic personalized interaction system and method based on AI context awareness. It solves the problems that existing screen layout technologies cannot effectively balance the conflict between users' personalized needs and multiple business objectives (such as inventory and revenue) within a limited space, as well as the lack of proactive detection mechanisms in scenarios where user intent is ambiguous, and the inability to transform abstract asset purchasing power into intuitive visual perception to lower the decision-making threshold.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an AI-based context-aware dynamic personalized interaction system, the system comprising: The multimodal context awareness module is configured to collect user behavior flow data, asset status data, and environmental context data in real time, and map the behavior flow data, asset status data, and environmental context data to a unified feature space to generate a context feature vector. The intelligent decision-making and interaction engine module is communicatively connected to the multimodal context perception module. It is configured to receive the context feature vector, identify the clarity of the user's intent based on the confidence difference mechanism, and adaptively switch between the precise adaptation mode and the intent detection mode accordingly. In the precise adaptation mode, it determines the allocation scheme of screen resources based on the multi-agent game mechanism, and performs nonlinear mapping of the visual attributes of the components in combination with the points purchasing power ratio to generate a UI configuration strategy that includes a layout structure list and a set of visual parameters. The dynamic rendering execution module is deployed on the client side and communicates with the intelligent decision and interaction engine module. It is configured to receive the UI configuration strategy, parse atomic components based on the server-driven UI architecture, and use affine transformation matrix and fragment shader parameters to transform the UI configuration strategy into a visual graphical user interface in real time. It is also responsible for detecting the structured emission of component interaction signals. The feedback optimization closed-loop module is communicatively connected to the dynamic rendering execution module and the intelligent decision-making and interaction engine module. It is configured to track the post-interaction data of the entire link, calculate the multi-dimensional business composite reward value, update the agent's policy network parameters through an online learning mechanism, and calibrate the intent recognition model using strongly supervised samples.

[0007] A second aspect of this invention provides a dynamic personalized interaction method based on AI context awareness, the method comprising the following steps: Step S100: Collect user behavior flow data, asset status data and environmental context data in real time through the multimodal scenario awareness module, and map the behavior flow data, asset status data and environmental context data to a unified feature space to generate scenario feature vectors; Step S200: Receive the scenario feature vector through the intelligent decision-making and interaction engine module, identify the clarity of the user's intent based on the confidence difference mechanism, and adaptively switch between the precise adaptation mode and the intent detection mode accordingly. Step S300: In the precise adaptation mode, the intelligent decision-making and interaction engine module determines the allocation scheme of screen resources based on the multi-agent game mechanism and generates a macro layout strategy containing a layout structure list. Step S400: The intelligent decision-making and interaction engine module combines the points purchasing power ratio to perform non-linear mapping on the visual attributes of the components, generating a micro-UI configuration strategy containing a set of visual parameters. Step S500: Receive the macro layout strategy and the micro UI configuration strategy through the dynamic rendering execution module, parse atomic components based on the server-driven UI architecture, render a visual graphical user interface in real time using the affine transformation matrix and fragment shader parameters, and be responsible for detecting the structured emission of component interaction signals. Step S600: Track the post-hoc data of the entire interaction through the feedback optimization closed-loop module, calculate the multi-dimensional business composite reward value, update the agent's policy network parameters through the online learning mechanism, and calibrate the intent recognition model using strongly supervised samples.

[0008] This invention provides a dynamic personalized interaction system and method based on AI context awareness. It has the following beneficial effects: 1. This invention intelligently identifies the clarity of user intent through a dual-mode adaptive switching mechanism based on confidence difference: when confidence is high, a precise adaptation mode is used to directly generate the layout; when confidence is low, it automatically switches to intent detection mode to obtain more information. This solves the problem of decreased user experience caused by forced matching in traditional recommendation systems during cold starts or weak intent scenarios, achieving a dynamic balance between interaction efficiency and intent recognition accuracy.

[0009] 2. This invention introduces agents representing different objectives such as points consumption, inventory clearance, and interest retention into a virtual bidding process through a screen resource allocation mechanism based on multi-agent game theory. This overcomes the limitations of traditional static templates or single rules that cannot accommodate multiple business conflicts, achieving a Nash equilibrium between personalized user experience and business objectives such as inventory turnover and cash flow within a limited screen space, thus maximizing the comprehensive business value of traffic.

[0010] 3. This invention establishes a nonlinear mapping between purchasing power ratio and visual attributes, utilizing a mixture of S-shaped step and Gaussian peak functions to transform the abstract relationship between points balance and product price into intuitive component size, transparency, and dynamic lighting effects. Users can intuitively perceive the ease of obtaining a product without complex mental calculations, enhancing the visual stimulation of items nearing redemption, lowering the decision-making threshold for users, and increasing conversion rates.

[0011] 4. This invention optimizes the closed loop through online feedback based on streaming data, and uses a temporal difference algorithm to update the agent's policy parameters in real time according to actual interaction results, and dynamically calibrates the determination threshold for intent recognition. This enables the system to have self-evolution and correction capabilities, adapting to constantly changing user behavior patterns and business environments, without relying on high-frequency offline model retraining, ensuring the robustness and timeliness of the decision-making strategy. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the data processing flow of the multimodal scene perception module of the present invention; Figure 4 This is a schematic diagram of the processing flow of the intent recognition and interaction mode triage module of the present invention; Figure 5 This is a schematic diagram of the screen layout generation process based on multi-agent game theory according to the present invention. Figure 6 This is a schematic diagram of the operation process of the feedback optimization closed-loop module of the present invention. Detailed Implementation

[0013] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Please see the appendix Figure 1 , Figure 1 This is a schematic diagram of a dynamic screen layout generation system based on multi-agent game theory according to an embodiment of the present invention. The present invention provides a dynamic screen layout generation system based on multi-agent game theory, comprising: The multimodal context awareness module is configured to collect heterogeneous user behavior data, asset status data, and environmental context data in real time, and perform feature cleaning, encoding, and fusion operations to transform discrete raw signals into continuous high-dimensional context feature vectors.

[0015] The intelligent decision-making and interaction engine module is communicatively connected to the multimodal context awareness module. It is configured to receive the context feature vector, identify the clarity of the user's intent based on a confidence differential mechanism, and adaptively switch between a precise adaptation mode and an intent detection mode accordingly. In precise adaptation mode, the intelligent decision-making and interaction engine module determines the allocation scheme of screen resources through a multi-agent bidding game mechanism and performs a non-linear mapping of the visual attributes of components based on the points purchasing power ratio, generating a UI configuration strategy that includes a layout structure list and a set of visual parameters.

[0016] The dynamic rendering execution module is deployed on the client side and communicates with the intelligent decision-making and interaction engine module. It is configured to receive the UI configuration strategy, parse atomic components based on the server-driven UI architecture, and use affine transformation matrix and fragment shader parameters to transform the configuration strategy into a visual graphical user interface in real time. It is also responsible for detecting the structured emission of component interaction signals.

[0017] The feedback optimization closed-loop module is communicatively connected to the dynamic rendering execution module and the intelligent decision-making and interaction engine module. It is configured to track the post-interaction data of the entire link, calculate the multi-dimensional business composite reward value, update the agent's policy network parameters using the temporal difference algorithm through an online learning mechanism, and calibrate the intent recognition model using strongly supervised samples.

[0018] Please see the appendix Figure 2 , Figure 2 This is a flowchart illustrating a dynamic screen layout generation method based on multi-agent game theory according to an embodiment of the present invention. The present invention provides a dynamic screen layout generation method based on multi-agent game theory, comprising the following steps: S100: Captures raw signals across the entire domain through a multimodal scene awareness module, performs nonlinear feature mapping on asset data, performs deep semantic encoding on behavior sequences, and constructs a unified multimodal feature vector space.

[0019] S200: It performs probability distribution reasoning in the intent space through the intelligent decision-making and interaction engine module, determines the interaction mode based on confidence difference, and selects the detection component that can maximize information gain when the intent is ambiguous.

[0020] S300: Through the intelligent decision-making and interaction engine module, it performs resource discretization of the screen space, instantiates a set of business intelligent agents, calculates virtual bids based on a hybrid value prediction model, and performs global conflict resolution and winner-take-all arbitration to generate a layout structure.

[0021] S400: Calculates the points purchasing power ratio through the intelligent decision-making and interaction engine module, activates visual saliency based on the blending function, and dynamically maps the visual saliency to rendering parameters such as component size, transparency, and lighting effects.

[0022] S500: The dynamic rendering execution module recursively parses and reconstructs the view tree based on the atomic component library for the issued configuration strategy, calculates the visual transformation matrix and applies styles, and dynamically binds interactive events.

[0023] S600: Through feedback optimization of the closed-loop module, it tracks the posterior data of the interaction, calculates the multi-dimensional business reward function, updates the bidding strategy network parameters based on time-series difference, and adaptively calibrates the confidence threshold of the intent recognition model.

[0024] The following description, in conjunction with the accompanying drawings, details the dynamic screen layout generation system based on multi-agent game theory according to an embodiment of the present invention.

[0025] The multimodal context-aware module, acting as the system's perception front-end, is responsible for constructing a digital holographic mapping of the user's current state. This module acquires behavioral flow data, asset status data, and environmental context data through parallel acquisition channels. For behavioral flow data, the module not only records click and browsing events but also structurally defines micro-features such as touch action types, dwell time, and swipe speed. For asset data, the module synchronizes points balance and expiration time via an API interface. For environmental data, the module reads network and device status.

[0026] The multimodal context awareness module integrates a feature preprocessing unit and a feature fusion unit. For numerical asset data, this module does not directly use the raw values ​​but instead constructs a feature vector representing the urgency of the assets. Through logarithmic smoothing and ratio normalization, the absolute values ​​are transformed into relative weights reflecting purchasing power and asset stress coefficients. For behavioral sequences, the module utilizes an attention mechanism to capture long-range dependencies. Finally, the features from each dimension are concatenated and stitched together into a unified context feature vector, which is then transmitted to downstream modules.

[0027] The intelligent decision-making and interaction engine module, as the core processing hub of the system, undertakes the dual functions of intent recognition and resource allocation. This module first performs intent reasoning on the input scenario feature vector, calculating the conditional probability distribution in the discrete intent space. Instead of directly accepting the highest probability result, this module calculates the confidence difference value of the probability distribution.

[0028] When the confidence level is lower than the preset threshold, the intelligent decision-making and interaction engine module enters the intent detection mode. By calculating the discrimination function of candidate components, it selects detection components that can eliminate intent ambiguity and displays them first. When the confidence level meets the requirements, it enters the precise adaptation mode and starts the multi-agent game process.

[0029] In the game-theoretic process, the intelligent decision-making and interaction engine module divides the screen into ordered discrete resource slots and instantiates business agents representing different optimization objectives such as points consumption, inventory clearance, and interest retention. Each agent calculates a virtual bid for specific slot content based on the current scenario and its own utility function. This bid integrates the estimated click-through rate and the inherent value of the business. The system arbitrates the bid matrix through a winner-takes-all strategy, generating a layout structure list that maximizes overall business utility.

[0030] Subsequently, the intelligent decision-making and interaction engine module further refines the control of visual attributes. This module calculates the purchasing power ratio between the product price and the user's assets, and uses a non-linear activation strategy combining a sigmoid function and a Gaussian function to calculate the visual saliency coefficient of the component. This coefficient is mapped to specific size scaling ratios, transparency, and dynamic lighting effects, and encapsulated within the UI configuration strategy.

[0031] The dynamic rendering execution module, acting as the client-side execution unit, pre-configures a component library based on atomic design theory. This module receives and parses the UI configuration strategy, constructing a view tree through depth-first traversal. It utilizes an affine transformation matrix to convert server-side visual parameters into graphics rendering instructions, enabling dynamic component deformation and effects loading. Simultaneously, this module can identify and probe components, intercept default navigation behaviors, and generate structured signals carrying confirmation intent to send back to the server.

[0032] The feedback optimization closed-loop module is responsible for the system's adaptive evolution. This module tracks user interaction behavior after each layout display, including explicit clicks and implicit dwell times. Based on the type of the winning agent, it calculates the actual business reward value using weighted composite logic. Based on this reward value, the module uses reinforcement learning algorithms to calculate gradients and update the valuation network parameters of each agent, making their virtual bids approximate the real business value. Simultaneously, the module utilizes feedback data from the probing mode to construct strongly supervised samples, fine-tuning the intent recognition model and dynamically adjusting the confidence threshold, enabling online iteration of the system strategy.

[0033] Please see the appendix Figure 3 , Figure 3This is a schematic diagram of the data processing flow of a multimodal scene perception module according to an embodiment of the present invention. The multimodal scene perception module, as the input end of the system, is configured with a parallel data processing pipeline for mapping heterogeneous user behavior data and static asset attribute data to a unified feature space. The multimodal scene perception module internally includes a real-time signal acquisition unit, a feature preprocessing unit, a sequence feature encoding unit, and a feature fusion unit. These units collaboratively execute the following data processing steps: S110: Real-time capture and structured definition of raw signals across the entire domain.

[0034] The real-time signal acquisition unit is configured with three independent data channels, used to capture behavior flow data, asset status data, and environmental context data, respectively. In the behavior flow data acquisition channel, the real-time signal acquisition unit not only records regular page browsing and click events, but is also configured to capture micro-touch features that characterize user decision-making psychology. This real-time signal acquisition unit will collect the first... Interaction events A structured definition is a multidimensional feature tuple: ; In the formula, The first one collected by the system A structured feature tuple of each interactive event; An enumeration value is used to identify the type of touch action, which is used to distinguish discrete operation behavior categories such as single click, long press, quick swipe or hover. It serves as a unique resource identifier for the interacting object, used to precisely locate the specific UI element or product component ID that is manipulated by the user within the system. The effective dwell time of a user's gaze or touch focus on an interactive object, usually measured in milliseconds, is used to characterize the depth of the user's attention to that object; This is a scalar value representing the pixel acceleration of screen swiping during interaction, used to characterize the user's eagerness or information filtering rate while browsing.

[0035] In the asset status data acquisition channel, the real-time signal acquisition unit synchronizes the user's points financial attributes via an API interface, including the total points balance. Points quota that are about to expire and the remaining time window before expiration. .

[0036] In the environmental context data acquisition channel, the real-time signal acquisition unit reads the physical status parameters of the terminal device, including the network connection type identifier, the device power percentage, and the current system timestamp.

[0037] S120: Nonlinear feature mapping for asset data. The feature preprocessing unit is configured to standardize heterogeneous data. For numerical asset status data, this feature preprocessing unit does not directly use the original values, but instead constructs an intermediate vector of asset features representing the urgency of the assets. .

[0038] This feature preprocessing unit calculates the intermediate vector of asset features using a combination of log-smoothing and ratio normalization. The calculation formula is as follows: ; In the formula, This is an intermediate vector of asset features after nonlinear mapping, used to represent the user's current asset urgency status; This is the absolute value of the total balance of points currently held in the user's account; The amount of points in a user's account that are marked as expiring soon; It is a preset non-zero smoothing correction constant used to prevent numerical anomalies caused by the total balance being zero, resulting in a zero denominator in division operations. This represents the remaining time window before points expire, measured in standard time units (such as days or hours). This is the numerical bias constant.

[0039] Using this formula, the feature preprocessing unit transforms absolute integral values ​​into relative weighted features that reflect the specific scenario of a user's points expiring soon. For categorical environmental data, the feature preprocessing unit uses one-hot encoding to map it into sparse environmental vectors. .

[0040] S130: Deep Semantic Encoding of Behavioral Sequences. The Sequence Feature Encoding Unit is configured to extract long-range dependencies and implicit intents from user behavior sequences. This Sequence Feature Encoding Unit is built on a Transformer network architecture, and its input is the sequence of interaction events generated in the preceding steps. ,in The sequence length is given.

[0041] The sequence feature encoding unit first uses an embedding layer to encode discrete interaction events. The sequence feature encoding unit is mapped to a low-dimensional dense vector and then superimposed with a positional encoding vector to preserve temporal information. Subsequently, this sequence feature encoding unit utilizes a multi-head self-attention mechanism to calculate the semantic association strength between events within the sequence. The multi-head self-attention mechanism captures the user's interest migration path from browsing to following by parallelly computing the scaled dot product attention between the query matrix, key matrix, and value matrix. After multiple layers of attention convergence and average pooling operations, this sequence feature encoding unit outputs a behavioral sequence feature vector representing the user's current dynamic interests. .

[0042] S140: Construction and Fusion of Multimodal Feature Vector Space. The feature fusion unit is configured to integrate features from the above dimensions to construct a unified scenario description for downstream decision-making. This feature fusion unit first integrates the intermediate vectors of asset features... The input fully connected layer is used for dimensional projection to obtain the asset feature vector. Simultaneously, the sparse environment vector Mapped to environmental feature vectors through the embedding layer .

[0043] Finally, the feature fusion unit performs a vector concatenation operation to generate a fused context feature vector. : ; In the formula, To step at the current time The generated fusion scenario feature vector serves as high-dimensional input data for the downstream intelligent decision engine. It is a deep behavioral sequence feature vector output by the sequence feature encoding unit, representing the evolution of the user's dynamic interests; For the intermediate vector of asset characteristics The asset feature vector generated after dimensional projection through the fully connected layer; This is an environmental feature vector generated from the original environmental data after one-hot encoding and embedding layer mapping; This is a vector concatenation operator.

[0044] The feature vector of this scenario As a dense vector containing user micro-behavioral intentions, macro-asset pressures, and external environmental constraints, it is transmitted to the intelligent decision-making and interaction engine module.

[0045] Please see the appendix Figure 4 , Figure 4 This is a schematic diagram of the processing flow of an intent recognition and interaction pattern triage module according to an embodiment of the present invention. The intelligent decision-making and interaction engine module executes a dynamic decision-making mechanism through this sub-flow. The intelligent decision-making and interaction engine module internally includes an intent reasoning unit, a confidence triage control unit, and a detection strategy generation unit. These units collaboratively execute the following processing steps: S210: Probability distribution reasoning of the intent space. The intent reasoning unit stores a predefined discrete intent space. ,in This represents the total number of intent categories. The intent reasoning unit is equipped with a classification network built on a multilayer perceptron, used to transmit the context feature vectors transmitted from the multimodal context awareness module. Mapped to this intent space.

[0046] The intent reasoning unit first utilizes a multi-layer fully connected network to process the feature vector. High-order feature extraction is performed, followed by the calculation of the conditional probability distribution of the current scenario belonging to each intent category using the Softmax normalization function. Here, the output is defined as... The predicted probability for each intent category is: This probability value represents the likelihood that a specific business objective is implied behind a user's current behavior.

[0047] S220: Pattern triage decision based on confidence level difference. The confidence level triage control unit is configured to assess the determinism of the aforementioned probability distribution and switch the system's interaction mode accordingly. This confidence level triage control unit does not directly accept the intent corresponding to the highest probability, but instead calculates a confidence index of the prediction result. .

[0048] In this embodiment, the confidence level split control unit uses the primary and secondary probability difference method to calculate the confidence level. This is the difference between the highest probability value and the second highest probability value, calculated using the following formula: ; In the formula, This is a confidence index for predicting current intent, used to quantify the degree of dominance of the intent. The maximum probability value in the intent probability distribution corresponds to the most likely intent category in the current prediction; This is the second-largest probability value in the intention probability distribution, corresponding to the maximum value among the remaining intentions after excluding the highest probability intention; The conditional probability function output by the intention reasoning unit; and These are indices for different intent categories in the intent space; The context feature vector input at the current time step.

[0049] This formula is used to quantify the dominance of an intention when there are two or more intentions with similar probabilities. The value decreased significantly.

[0050] Subsequently, the confidence-based diversion control unit will calculate the... Compared with the preset judgment threshold Comparison is performed by executing the traffic splitting strategy function. Determine the current system status: ; In the formula, The output state of the interaction mode routing strategy function determines the subsequent execution branch of the system; The status indicator for accurate adaptation mode indicates that the system has entered the direct layout generation process with conversion efficiency as the goal; This serves as a status indicator for the intent detection mode, signifying that the system has entered the process of generating detection components with the goal of information acquisition. The confidence index is calculated using the aforementioned formula; The preset or dynamically adjusted diversion threshold serves as a boundary value to distinguish between clear and ambiguous intentions.

[0051] When the judgment result is In (precise adaptation mode), the confidence-based traffic control unit locks the intent corresponding to the highest probability as the deterministic target and triggers the subsequent layout generation process; when the determination result is... In the (intent detection mode), the confidence diversion control unit activates the detection strategy generation unit.

[0052] S230: Optimal selection of detection components that maximize information gain. The detection strategy generation unit, in response to the activation signal of the intent detection mode, is responsible for selecting detection components from a pre-set component library that can eliminate intent ambiguity. This detection strategy generation unit first identifies the subset of intents that cause confusion based on a probability distribution. .

[0053] The detection strategy generation unit calculates a candidate component library based on the principle of maximizing information distinguishability. The scores of each component. For any candidate component The selection strategy formula is as follows: ; In the formula, The selected optimal detection components will serve as the core elements of the subsequent interactive interface; The system has a pre-built candidate library of detection components, which includes a variety of component templates with high interaction damping properties. This refers to any probe component object that is traversed in the candidate library; Operators that take the independent variable that maximizes the objective function; This is a set of confusing intentions that lead to ambiguous intent judgment, selected based on the current probability distribution. To obfuscate any specific intent element in the intent set; For the purpose In the current context The conditional probability value is used as the weighting coefficient; The discrimination scoring function value represents the component. Interaction options and specific intents The strength of semantic association or distinguishing ability between them.

[0054] This formula ensures that the components selected by the system can, to the greatest extent possible, distinguish the current obfuscating intent through a single user interaction.

[0055] S240: Feedback embedding and state update of the probe component. After selecting the probe component, the probe strategy generation unit instantiates and configures it, defining it as an interface element with high interaction damping properties. This probe strategy generation unit embeds structured feedback markers in the component's interaction interface. When the probe component is triggered by the user, the system does not perform a conventional page jump, but instead generates a synthetic event with a clear intent label and sends it back to the input, forcibly refreshing the context feature vector. This increases the confidence level of the next inference. Exceeding the threshold This enables closed-loop switching of interaction modes.

[0056] Please see the appendix Figure 5 , Figure 5 This is a schematic diagram of a screen layout generation process based on multi-agent game theory according to an embodiment of the present invention. To resolve conflicts in the allocation of multi-objective resources within a limited screen space, the intelligent decision-making and interaction engine module is configured with a screen space discretization unit, a business agent collaboration unit, a virtual bidding calculation unit, and a layout arbitration unit. Each unit collaboratively executes the following processing steps: S310: Screen space resource discretization definition. The screen space discretization unit is configured to map the physical display area of ​​a terminal device to resource slots that can be quantified and priced. This screen space discretization unit divides the screen into an ordered set of discrete slots. Each slot This corresponds to a container node in the final rendered view tree.

[0057] Based on ergonomic principles, the screen space discretization unit assigns a basic visibility weight to each slot, including the top slot of the first screen. The slot with the highest weight has its weight decreasing non-linearly as it moves down the position. This discretization transforms the continuous layout problem into a discrete optimization problem for a fixed resource location.

[0058] S320: Instantiation and Utility Configuration of Business Agents. The business agent collaboration unit is configured with multiple independently running agent instances, forming a bidding subject set. Each agent instance is encapsulated as an optimizer for specific business metrics.

[0059] In this embodiment, the set of agents includes at least: an agent that consumes points. Configured to prioritize high-value items to maximize asset consumption rate; Inventory depletion intelligent agent Configured to prioritize products with high inventory-to-sales ratios to accelerate inventory turnover; interest-based retention intelligent agent. The configuration is used to recommend content that matches the user's historical preferences to improve click-through rates. Each agent maintains an independent pool of candidate content.

[0060] S330: Virtual bidding calculation based on a hybrid value model. The virtual bidding calculation unit is configured to calculate the bidding intentions of each agent for a specific slot. For any slot to be filled... This virtual bidding computation unit drives each intelligent agent. Based on the current scenario feature vector its candidate content Conduct a value assessment.

[0061] The virtual bidding calculation unit uses a hybrid value model to calculate the virtual bid. This model combines the probability of traffic monetization with the inherent value of the business, and the calculation formula is as follows: ; In the formula, For the first The business intelligence agent targets the first Each screen resource slot is used for candidate content. The calculated virtual value; For the first The risk preference balance coefficient of each business agent, with a value range of [0,1], is used to adjust the agent's emphasis on traffic efficiency and inherent business value. The conditional probability output by the click-through rate prediction model represents the feature vector in the current context. Next, the candidate content will be... Displayed in slot When a user makes a click. The estimated probability; For the current time step Multimodal scenario feature vectors; For specific content items in the candidate content pool of the intelligent agent; The first screen space discretized One resource slot; For the first Business Intelligent Agent Candidate content is categorized based on its specific business logic (such as inventory depth, points unit price). The normalized business utility value obtained after evaluation; For the first to participate in the bidding A business intelligence agent instance. Through adjustment... The system can achieve a smooth transition between traffic-driven and revenue-driven approaches.

[0062] S340: Global Conflict Resolution and Winner-Win Arbitration. The layout arbitration unit is configured to determine the final screen content structure based on the bidding results. This layout arbitration unit receives the bidding matrix from all agents for all slots and executes a greedy arbitration strategy with deduplication constraints.

[0063] For each slot, the layout arbitration unit selects the agent with the highest bid and its recommended content as the winner. To ensure the diversity of interface content, the layout arbitration unit maintains a set of selected content. And during subsequent arbitration of slots, those already existing will be forcibly removed. The candidate content is selected. Finally, the layout arbitration unit outputs a list of layout structures containing component IDs, types, and their order. This list represents the Nash equilibrium state of achieving multiple business objectives in the current user scenario.

[0064] After completing the macro-level screen layout planning, the intelligent decision-making and interaction engine module utilizes this sub-process to perform fine-tuning of the micro-level visual presentation. This module internally includes a purchasing power ratio calculation unit, a visual salience activation unit, and a rendering parameter injection unit. Each unit establishes a physical mapping between the financial attributes of the points system and the UI pixel attributes, and collaboratively executes the following processing steps: S410: Quantitative calculation of the points-based purchasing power ratio. The purchasing power ratio calculation unit is configured to traverse the layout structure list and identify product components with transaction attributes. For the first item in the list... Each product component, the purchasing power ratio calculation unit obtains its points redemption price. And combined with the user's current points balance Conduct an asset coverage assessment.

[0065] The purchasing power ratio calculation unit uses a smoothing correction model to calculate the purchasing power ratio. The calculation formula is as follows: ; In the formula, For the first The points-purchasing power ratio of each product component is used to quantify the user's ability to cover the product price with their assets. The user's current real-time points account balance is obtained from the multimodal context awareness module; This refers to the user's current additional rights limit that can be used to redeem this product or the credit limit that allows overdraft; For the first The current points redemption price for each product component; This is a pre-defined non-zero smoothing correction term (small constant) used to prevent calculation anomalies caused by a zero denominator due to a zero commodity price.

[0066] Calculated A dimensionless numerical value used to represent the relative relationship between user assets and commodity prices: when When <1, it indicates insufficient integration; when When ≈1, it represents the critical exchange state; when A value of 1 indicates an abundance of assets.

[0067] S420: Visual saliency nonlinear activation based on a mixture function. The visual saliency activation unit is configured to transform the aforementioned linear purchasing power ratio into a nonlinear visual saliency coefficient. This visual saliency activation unit abandons the linear visual scaling logic and adopts a hybrid activation strategy that combines an S-shaped step function and a Gaussian peak function to highlight critical exchange and high-value products.

[0068] The visual saliency activation unit is based on the following nonlinear activation function. Calculate the visual saliency coefficient: ; In the formula, For the first The visual saliency coefficient of each product component, which is directly mapped to the size, transparency and special effects parameters during UI rendering; This is a preset basic visual weight constant for the component, used to ensure that the component has a minimum level of visibility; The step gain amplitude parameter determines the upper limit of the visual weight increase when the purchasing power ratio meets the exchange conditions; is the base of the natural logarithm; The slope parameter of the Sigmoid function is used to control the steepness of the step change in visual weight as the purchasing power ratio changes; The first step calculated by the preceding steps The purchasing power ratio of points for each product component; The peak gain height parameter of the Gaussian function is used to define the intensity of the additional visual reward imparted at the critical exchange point; The width parameter of the Gaussian function controls the rate of peak gain decay near the critical point, i.e., it controls the sensitive window range for critical state recognition. This function ensures that items that just deplete the user's balance receive the highest visual weight.

[0069] S430: Dynamic mapping and injection of visual rendering parameters. The rendering parameter injection unit configuration is used to map abstract visual saliency coefficients. Transform into a specific set of UI configuration parameters The rendering parameter injection unit has a pre-defined set of parameter mapping rules, covering geometric attributes, optical attributes, and dynamic interactive attributes.

[0070] In geometry attribute mapping, this render parameter injection unit will scale the component. Set as The positive correlation function causes highly salient items to occupy a larger pixel area on the interface. In optical attribute mapping, this rendering parameter injection unit is based on... The opacity and color saturation are dynamically adjusted based on the range of values. For For non-redeemable items below a preset threshold, the rendering parameter injection unit reduces their opacity parameter to create visual suppression. In dynamic interactive attribute mapping, this rendering parameter injection unit introduces an effect activation threshold. When calculated At that time, the rendering parameter injection unit is in the parameter set The system writes dynamic lighting effect markers to instruct the client to load breathing light or flowing light effects. Finally, this rendering parameter injection unit encapsulates the complete parameter set into the layout structure list, generating a UI configuration strategy that can be parsed by the client.

[0071] The dynamic rendering execution module, acting as the system's client-side execution terminal, is configured with a server-driven user interface rendering engine to transform abstract configuration strategies into a visual graphical interface. Internally, this module integrates an atomic component management unit, a strategy recursive parsing unit, a visual transformation rendering unit, and an interactive signal processing unit. These units collaboratively execute the following processing steps: S510: Declarative definition of the atomic component library. The atomic component management unit maintains a pre-built, standardized atomic component library, constructed based on atomic design theory, which decouples interface elements into minimal reusable units. Each component in this unit... None of them contain hard-coded style logic; instead, they expose a set of dynamically configurable property parameters through a declarative interface. Attribute parameters At least including: basic content parameters Used to define image resource addresses and text data; layout constraint parameters This is used to define width and height size limits and flexible layout weights; as well as visual rendering parameters. This is used to define the transparency channel, scaling factor, and border lighting effects. This atomic component management unit uses the aforementioned interface to decouple the front-end rendering layer from the business logic layer.

[0072] S520: Recursive parsing and view reconstruction of configuration strategies. The strategy recursive parsing unit is configured to receive UI configuration strategies from the server. These strategies are transmitted in a tree data structure, containing a list of layout structures. and the corresponding parameter set.

[0073] This strategy recursive parsing unit incorporates depth-first traversal logic, configured to recursively parse the configuration strategy. For each node in the tree structure, the unit indexes the corresponding component class in the atomic component library based on the type identifier and instantiates it as a virtual view object. During this process, the unit establishes a view hierarchy based on node attributes, constructing a view tree that maps to the server-side game result.

[0074] S530: Visual parametric rendering based on affine transformation. The visual transformation rendering unit is configured to traverse the view tree, mapping parameter sets to transformation instructions from the underlying graphics rendering pipeline. For component geometry adjustments, this visual transformation rendering unit utilizes affine transformation matrices to achieve dynamic scaling.

[0075] Let the original geometric coordinate matrix of the component be... The scaling parameters sent by the server are The visual transformation rendering unit constructs the scaling transformation matrix. Its mathematical expression is as follows: Based on this matrix, the visual transformation rendering unit calculates the final rendering coordinate matrix. : ; In the formula, The target geometric coordinate matrix for the component to be drawn in the graphics rendering pipeline. This matrix contains the absolute position data of the component vertices in the screen pixel coordinate system after transformation, which directly determines the coverage area of ​​the component in the final rendering. This is the initial geometric coordinate matrix generated during the component's instantiation phase, which is composed of the component's layout constraint parameters. The initially defined set of vertex coordinates constitutes the original geometric state without visual enhancement processing; This is an affine scaling transformation matrix constructed based on the visual saliency coefficient. This matrix is ​​a 3×3 linear transformation operator used to scale the scaling factor. Mapped to a two-dimensional graphics coordinate space, this drives the graphics processing unit to perform pixel-level linear scaling operations on component primitives.

[0076] For transparency and color saturation attributes, this graphics processing unit directly modifies the parameters of the fragment shader in the graphics pipeline. If the configuration parameters include dynamic lighting effects, this graphics processing unit will attach a separate animation layer on top of the component layer and load a preset dynamic texture resource.

[0077] S540: Signal interception and structured emission of interactive events. The interactive signal processing unit is configured to manage the dynamic interaction logic of components. This unit features component type recognition; when a probe component is detected, it automatically intercepts the default route jump event. The unit also includes a built-in signal transmitter configured to intercept user clicks on probe components. At the same time, a structured signal carrying explicit intent parameters is generated. The structure of this signal is defined as follows: ; In the formula, This is a structured intent feedback signal data packet constructed by the interactive signal processing unit after intercepting a click event at the underlying level. This is the identifier for the event type field in the signal data structure; A specific string constant value is assigned to the event type field to uniquely identify to the feedback optimization closed-loop module that the signal originates from a valid user response to the probe component, rather than a normal page redirection request. This is the identifier for the payload field in the signal data structure that carries core business data. This is the label value for the intent category explicitly selected by the user through touch interaction, which directly corresponds to the system's predefined intent space. A specific discrete intent element in the text represents the true user goal after disambiguation; This is the identifier for the timestamp field in the signal data structure that records the moment an event occurs. This provides the system kernel with a precise system time value in milliseconds when it captures user interaction operations (such as the moment a finger is lifted). This value is used for timing alignment and causal analysis during subsequent end-to-end data tracing.

[0078] This interactive signal processing unit will process structured signals. The feedback optimization closed-loop module is pushed in real time and triggers a partial redraw command on the client to update the display status of the relevant area without refreshing the entire page.

[0079] Please see the appendix Figure 6 , Figure 6 This is a schematic diagram of the operation flow of a feedback optimization closed-loop module according to an embodiment of the present invention. As a self-evolutionary correction device for the system, the feedback optimization closed-loop module does not rely on offline batch training, but is equipped with an online learning mechanism based on streaming data for continuously calibrating the policy parameters in the intelligent decision-making and interaction engine module. The feedback optimization closed-loop module integrates an interaction end-link tracking unit, a multi-dimensional reward composite calculation unit, a policy network parameter update unit, and an intent model adaptive calibration unit. These units collaboratively execute the following processing steps: S610: End-to-end tracing and status marking of post-interaction data. The end-to-end interaction tracing unit is configured to initiate tracing tasks after the interface is rendered. This end-to-end interaction tracing unit assigns a globally unique session tracing ID to each generated UI layout and is configured with a listening interface to capture subsequent user actions.

[0080] This interaction end-to-end tracing unit defines an interaction result vector. This is used to record explicit and implicit feedback signals. Interaction result vector. Includes binarized click signals , conversion signal and continuous dwell time .

[0081] For specific intelligent agents The acquired slot is used by the full-link interaction tracking unit to bind the user's feedback behavior with the agent's identifier, constructing a triplet sample containing contextual features, decision-making actions, and feedback results. And store it in the experience replay buffer.

[0082] S620: Composite Calculation of Multidimensional Business Reward Function. The multidimensional reward composite calculation unit is configured to quantitatively evaluate the business value of deployment strategies. Considering the heterogeneity of multiple business objectives in the system, this multidimensional reward composite calculation unit uses weighted composite logic to transform the interaction results into scalar reward values.

[0083] This multi-dimensional reward composite calculation unit is based on the winning agent. The type attribute dynamically calls the weight configuration, and the actual reward value is calculated according to the following formula. : ; In the formula, This is the actual composite reward value calculated based on the current user interaction behavior; , , , The reward weighting coefficient for the current winning agent type is used to adjust the proportion of clicks, conversions, retention, and negative feedback in the total reward; This is a binary click signal; it takes a value of 1 when a user clicks on the target area, and 0 otherwise. This represents the actual value of points earned by the user during the current interaction. If no redemption occurs, this value is 0. This represents the user's current total points balance, used to calculate the percentage of total assets consumed in this transaction. This is the hyperbolic tangent function, used to map time-dimension data to a standardized interval; The effective dwell time of a user in the current content display area; The preset baseline dwell time threshold is used to determine whether a user's dwell time constitutes effective attention; This is a negative feedback indicator. It is set to 1 when a user is detected to swipe across the current area at a speed exceeding a preset threshold, and 0 otherwise.

[0084] S630: Policy network parameter update based on temporal difference. The policy network parameter update unit is configured to minimize the deviation between the virtual bid and the actual business value. This policy network parameter update unit is constructed with a specific form of loss function. , used to measure the The estimated network parameters of each agent The accuracy.

[0085] The network parameter update unit of this strategy calculates the loss function using the mean squared error form. ; In the formula, For the first The loss function value of the network for each agent; For the first The current set of weight parameters of the evaluation network for each agent; This is a discount factor for future rewards, ranging from 0 to 1, used to adjust the current strategy's emphasis on long-term cumulative returns; This is the context feature vector for the next moment after the current interaction occurs; This refers to the bidding actions that each agent might take in the next moment; For the next moment of the scenario Take action below The estimated value function; This indicates that the action that maximizes the estimated value is selected from all available actions. For the first Each agent is based on the current parameters For the current situation and slots The given predictions yield value; This is the feature vector of the current situation.

[0086] The network parameter update unit of this strategy is equipped with a gradient descent optimizer, which updates the network parameters based on the calculated gradient. Perform parameter update operation: ; In the formula, The preset learning rate hyperparameter is used to control the step size of parameter updates; loss function Regarding network parameters The gradient vector; This represents an iterative update operation that assigns the calculated new parameter value to the original parameter.

[0087] Through this update mechanism, the strategy network parameter update unit automatically adjusts the weight parameters of each business agent in the bidding game based on actual data feedback.

[0088] S640: Adaptive calibration of the intent recognition model. The intent model adaptive calibration unit is configured to optimize the intent recognition model using feedback data from the probe component. When the system triggers the probe mode and the user produces a valid response, a confirmation intent is generated. At that time, the intention model adaptive calibration unit constructs strongly supervised samples to fine-tune the intention inference network.

[0089] Meanwhile, the intention model adaptive calibration unit is configured with a threshold dynamic adjuster to adjust based on the most recent time window. The confidence threshold is dynamically adjusted to improve the accuracy of intent prediction. The threshold adjustment formula is as follows: ; In the formula, The updated confidence threshold will be written back to the judgment logic of the intent recognition and interaction mode triage module, serving as the judgment benchmark for distinguishing between the accurate adaptation mode and the intent detection mode in the next system cycle. The confidence threshold used when the system actually performs the judgment at the current time step or in the previous update cycle; The preset adjustment step size coefficient is used to control the sensitivity of the threshold to dynamically respond to accuracy deviation, preventing the threshold from oscillating violently due to single statistical fluctuations. The most recent time window for system statistics The statistical value of the accuracy of intent prediction within the system refers to the percentage of samples in the probe mode where the user’s true intent confirmed through interactive feedback matches the intent with the highest initial prediction probability of the system. The preset target accuracy benchmark value for the system represents the expected baseline for intent recognition accuracy set by the system in the current business scenario to ensure user experience. When it is below this value, it will drive This reduces the frequency of triggering detection modes, thereby increasing the number of training samples collected.

[0090] This mechanism ensures that the system automatically balances the triggering frequency of the precise adaptation mode and the intent detection mode when the model maturity changes.

Claims

1. A dynamic personalized interaction system based on AI context awareness, characterized in that, The system includes: The multimodal context awareness module is configured to collect user behavior flow data, asset status data, and environmental context data in real time, and map the behavior flow data, asset status data, and environmental context data to a unified feature space to generate a context feature vector. The intelligent decision-making and interaction engine module is communicatively connected to the multimodal context perception module and is configured to receive the context feature vector, identify the clarity of the user's intent based on the confidence difference mechanism, and adaptively switch between the precise adaptation mode and the intent detection mode accordingly. In the precise adaptation mode, the allocation scheme of screen resources is determined based on the multi-agent game mechanism, and the visual attributes of components are non-linearly mapped by the points purchasing power ratio to generate a UI configuration strategy that includes a layout structure list and a set of visual parameters. The dynamic rendering execution module is deployed on the client side and communicates with the intelligent decision and interaction engine module. It is configured to receive the UI configuration strategy, parse atomic components based on the server-driven UI architecture, and use affine transformation matrix and fragment shader parameters to transform the UI configuration strategy into a visual graphical user interface in real time. It is also responsible for detecting the structured emission of component interaction signals. The feedback optimization closed-loop module is communicatively connected to the dynamic rendering execution module and the intelligent decision-making and interaction engine module. It is configured to track the post-interaction data of the entire link, calculate the multi-dimensional business composite reward value, update the agent's policy network parameters through an online learning mechanism, and calibrate the intent recognition model using strongly supervised samples.

2. The AI-based context-aware dynamic personalized interaction system according to claim 1, characterized in that, When generating the scenario feature vector, the multimodal scenario awareness module is configured to perform the following operations: Construct an intermediate vector of asset characteristics to represent the urgency of assets. The construction operation includes log smoothing the total balance of points to represent the purchasing power scale, calculating the proportion of near-expiry assets in total assets to represent the asset pressure coefficient, and calculating the inverse of urgency decay in the time dimension. The interaction event sequence in the behavior stream data is encoded using an attention-based neural network model to capture the semantic correlation strength between events within the sequence and generate a behavior sequence feature vector. The asset feature vector generated by dimensional projection of the intermediate asset feature vector, the behavior sequence feature vector, and the environmental feature vector encoded from the environmental context data are concatenated and spliced ​​to generate the scenario feature vector.

3. The AI-based context-aware dynamic personalized interaction system according to claim 1, characterized in that, The intelligent decision-making and interaction engine module is configured to identify the clarity of user intent based on a confidence differential mechanism in the following manner: The scenario feature vector is mapped to a predefined discrete intent space using a classification network, and the conditional probability distribution of the current scenario belonging to each intent category in the discrete intent space is calculated. The confidence index of the prediction result is calculated using the primary and secondary probability difference method, wherein the confidence index is the difference between the maximum probability value and the second largest probability value in the conditional probability distribution; The confidence index is compared with a preset judgment threshold. If the confidence index is less than the judgment threshold, it is determined that the intent detection mode is entered. If the confidence index is greater than or equal to the determination threshold, then it is determined that the system will enter the precise adaptation mode.

4. The AI-based context-aware dynamic personalized interaction system according to claim 3, characterized in that, In the intent detection mode, the intelligent decision-making and interaction engine module is also configured to: Identify the subset of intents that cause intent confusion, and filter probe components from a pre-built component library based on the principle of maximizing information distinguishability; The strategy for selecting detection components is as follows: traverse the candidate component library, calculate the score of each candidate component, the score is the weighted sum of the discrimination score function value between the interaction options of the candidate component and the specific obfuscation intent and the current probability value of the specific obfuscation intent, and select the candidate component with the highest score as the optimal detection component; Structured feedback markers are embedded in the interaction interface of the detection component. When the user triggers the detection component, a synthetic event with a clear intent label is generated and sent back to refresh the context feature vector.

5. The AI-based context-aware dynamic personalized interaction system according to claim 3, characterized in that, In the precise adaptation mode, the intelligent decision-making and interaction engine module is configured to determine the screen resource allocation scheme based on a multi-agent game mechanism in the following manner: Discretize the screen display area into ordered resource slots; Instantiate multiple business intelligence agents targeting different business metrics, wherein the business intelligence agents include at least a points consumption intelligence agent, an inventory clearance intelligence agent, and an interest retention intelligence agent; Each of the aforementioned business agents is driven to calculate a virtual bid for the resource slot to be filled. The virtual bid is calculated based on a hybrid value model, which is a weighted combination of the estimated click probability and the normalized business utility value. The weight of the weighted combination is adjusted by a risk preference balancing coefficient. A greedy arbitration strategy with deduplication constraints is executed to select the business agent with the highest virtual bid and the recommended content of the business agent as the winner for each resource slot, thereby generating the layout structure list.

6. The AI-based context-aware dynamic personalized interaction system according to claim 1, characterized in that, When performing non-linear mapping of the visual attributes of components, the intelligent decision-making and interaction engine module is configured to: Calculate the points purchasing power ratio of the product component, wherein the points purchasing power ratio is the ratio of the sum of the user's current points balance and the rights limit to the points redemption price of the product. A nonlinear activation strategy based on a mixture function is used to calculate the visual saliency coefficient. The mixture function is composed of a superposition of an S-shaped step function term and a Gaussian function term. The S-shaped step function term is used to generate a visual hierarchy leap when the integral purchasing power ratio crosses a critical point, and the Gaussian function term is used to form a local visual convexity when the integral purchasing power ratio is near the critical point. The visual saliency coefficient is dynamically mapped to a set of UI configuration parameters, including setting the component size scaling ratio as a positive correlation function of the visual saliency coefficient, and adjusting the component opacity and color saturation according to the range of the points purchasing power ratio.

7. The AI-based context-aware dynamic personalized interaction system according to claim 1, characterized in that, The dynamic rendering execution module is configured to: A pre-built atomic component library is maintained based on atomic design theory. The components in the atomic component library expose configurable attribute parameters through a declarative interface. The received UI configuration strategy is recursively parsed to construct a view tree, and the final rendering coordinate matrix of the component is calculated using the affine scaling transformation matrix constructed based on the visual saliency coefficient. Identify the probe components in the UI configuration strategy, intercept the default route jump events of the probe components, and generate a structured intent feedback signal data packet containing confirmation intent category labels and timestamps when a user click operation is detected.

8. The AI-based context-aware dynamic personalized interaction system according to claim 3, characterized in that, The feedback optimization closed-loop module is configured to: The weight configuration is dynamically invoked based on the type attribute of the winning agent, and the actual composite reward value is calculated. The composite reward value is a weighted sum of the click signal, the proportion of the value of the points exchange, the dwell time after normalization by the hyperbolic tangent function, and the negative feedback flag. A loss function is constructed to measure the accuracy of the agent's estimated network parameters. The loss function is calculated based on the mean squared error between the composite reward value, the estimated value of the next state, and the current virtual bid. The gradient descent optimizer updates the agent's policy network parameters based on the gradient of the loss function.

9. A dynamic personalized interaction system based on AI context awareness according to claim 8, characterized in that, The feedback optimization closed-loop module is also configured to perform adaptive calibration of the intent recognition model: When the system triggers the intent detection mode and the user generates a valid response, a strongly supervised sample containing confirmed intent is constructed to perform backpropagation fine-tuning on the intent reasoning network. The accuracy of intent prediction within the most recent time window is statistically analyzed, and the judgment threshold is adjusted using a threshold dynamic adjuster. The adjustment logic of the threshold dynamic adjuster is as follows: based on the deviation between the recent intention prediction accuracy and the preset target accuracy, and combined with the preset adjustment step size coefficient, the judgment threshold of the next system cycle is dynamically increased or decreased.

10. A dynamic personalized interaction method based on AI context awareness, characterized in that, Using the AI-based context-aware dynamic personalized interaction system according to any one of claims 1-9, the method includes the following steps: Step S100: Collect user behavior flow data, asset status data and environmental context data in real time through the multimodal scenario awareness module, and map the behavior flow data, asset status data and environmental context data to a unified feature space to generate scenario feature vectors; Step S200: Receive the scenario feature vector through the intelligent decision-making and interaction engine module, identify the clarity of the user's intent based on the confidence difference mechanism, and adaptively switch between the precise adaptation mode and the intent detection mode accordingly. Step S300: In the precise adaptation mode, the intelligent decision-making and interaction engine module determines the allocation scheme of screen resources based on the multi-agent game mechanism and generates a macro layout strategy containing a layout structure list. Step S400: The intelligent decision-making and interaction engine module combines the points purchasing power ratio to perform non-linear mapping on the visual attributes of the components, generating a micro-UI configuration strategy containing a set of visual parameters. Step S500: Receive the macro layout strategy and the micro UI configuration strategy through the dynamic rendering execution module, parse atomic components based on the server-driven UI architecture, render a visual graphical user interface in real time using the affine transformation matrix and fragment shader parameters, and be responsible for detecting the structured emission of component interaction signals. Step S600: Track the post-hoc data of the entire interaction through the feedback optimization closed-loop module, calculate the multi-dimensional business composite reward value, update the agent's policy network parameters through the online learning mechanism, and calibrate the intent recognition model using strongly supervised samples.

Images