Automatic building and monitoring method for information promotion account structure based on AI

Abstract

The invention discloses an AI-based information promotion account structure automatic construction and monitoring method. The method comprises the steps of S1, multi-dimensional user portrait construction and psychological account mapping; s2, automatic hierarchical design of a promotion account structure; s3, intelligent content matching and delivery strategy execution; and S4, adjusting and optimizing the structure of real-time monitoring and feedback driving. According to the method, the dynamic user portrait is constructed through the time sequence behavior encoder, the user is classified into different psychological account types such as'necessity budget 'and'emotional reward' by utilizing the machine learning model, the limitation of static labels is exceeded, and the dynamic user portrait is obtained by combining the dynamically updated'content feature-psychological account 'mapping strategy matrix. According to the method, the intrinsic psychological motivation and decision-making mode of the user can be accurately matched, so that the promotion content with stronger emotion resonance and more matched conversion motivation is generated or selected, and the attraction and conversion efficiency of the content are improved.

Description

Technical Field

[0001] This application belongs to the field of intelligent information processing technology, specifically involving an AI-based method for automating the construction and monitoring of information promotion account structures. Background Technology

[0002] With the rapid development of digital marketing and online advertising technologies, information promotion has become a core means for businesses to acquire users and enhance brand influence. Current mainstream promotion platforms (such as search engines, social media, and news feed applications) generally offer targeted advertising based on user tags. Advertisers or operations personnel typically need to manually build a multi-level, tree-structured promotion account in the backend to organize different ad campaigns, ad groups, and ad creatives, and set budgets, bids, audience targeting rules, etc. This process is highly dependent on manual experience, and as the user base expands, channels increase, and strategies become more complex, many bottlenecks are exposed.

[0003] At the user understanding level, existing technologies mostly segment audiences based on static, discrete demographic attributes or behavioral tags (such as age, gender, and recently clicked product categories). This segmentation method fails to deeply capture the underlying psychological motivations and dynamic contextual changes during user decision-making. For example, the mental accounts of the same user when rationally comparing prices to buy office supplies on a weekday are completely different from those when emotionally browsing luxury goods on a weekend. Existing systems often fail to identify and utilize these subtle but crucial differences, leading to mismatched content recommendations and low conversion rates.

[0004] At the account structure and strategy design level, there is currently a heavy reliance on manual operations and experience-based judgment by operations personnel. From designing account tiers based on promotional goals to allocating budgets for each node, selecting channels, and formulating bidding strategies, the entire process is time-consuming, labor-intensive, highly subjective, and difficult to scale. More importantly, this static structure cannot adapt to real-time changes in user distribution, market competition, and campaign performance, lacking an intelligent mechanism that can automatically explore the optimal structure and dynamically optimize it under constraints (such as total budget and channel limits).

[0005] While personalized recommendations are widely used in content matching and delivery, they are often disconnected from account structure management and bidding strategy optimization. The management of content creative libraries, the assembly and generation of personalized content, and the calculation of optimal bids and frequency control in a real-time bidding environment are often handled by different modules or manually, lacking a closed-loop system driven by a unified user mental model to achieve intelligent decision-making and execution across the entire process.

[0006] In summary, existing information promotion technologies suffer from a series of problems, including superficial user psychology insights, rigid and reliant on manual account structure building, a disconnect between content strategy and execution, a lack of automated and intelligent closed-loop system monitoring and optimization, and insufficient privacy compliance protection. Therefore, there is an urgent need for an automated and systematic solution that deeply integrates artificial intelligence technology to achieve everything from deep user understanding and intelligent structure building to precise content execution, real-time adaptive optimization, and embedded privacy protection mechanisms. Summary of the Invention

[0007] This application provides an AI-based method for automating the construction and monitoring of information promotion account structures, aiming to solve the problems that existing technologies struggle to deeply capture the intrinsic psychological motivations and dynamic situational changes during user decision-making, lack intelligent mechanisms for dynamic optimization, and are disconnected from personalized recommendations, account structure management, bidding strategy optimization, and other related processes.

[0008] An AI-based method for automating the construction and monitoring of information promotion account structures, the method comprising:

[0009] S1. Multi-dimensional user profile construction and mental account mapping: Collect user behavior data, preference characteristics and historical interaction records, and construct dynamic user profiles through time-series behavior encoders; Based on the dynamic user profiles, use machine learning models to determine the mental account type of the user, and establish a three-layer mapping relationship between users, mental account types and promotional content characteristics.

[0010] S2. Automated hierarchical design of promotion account structure: Based on the promotion goals, resource constraints and the user mental account hierarchical results obtained in S1, a hierarchical reinforcement learning model under constraints is used to automatically generate a multi-level promotion account structure, in which each account node encapsulates specific content strategies, delivery channels, resource budgets and target audience selection rules based on user profiles and mental account weights.

[0011] S3. Intelligent content matching and delivery strategy execution: Based on the user's real-time mental account type obtained in S1 and the strategy defined by the account node in S2, natural language processing and image recognition technology are used to match or generate personalized promotional content, and the delivery strategy is dynamically optimized in combination with a two-sided market model, and the delivery is executed through the channels bound to the account node in S2.

[0012] S4. Real-time monitoring and feedback-driven structural optimization: Real-time collection of performance indicator streams of each node in the account structure generated in S2, identification of anomalies through an integrated anomaly detection model, and generation of optimization actions in combination with the contextual multi-armed gambling machine model to dynamically adjust the account structure, resource allocation or content strategy.

[0013] Optionally, in S1, the temporal behavior encoder is a recurrent neural network based on an attention mechanism, used to model user behavior sequences to capture the evolution trend and context dependence of behavior patterns; the three-layer mapping relationship is specifically a mapping from user to mental account type, and a mapping from mental account type to promotional content features, wherein the promotional content features include information frames, price anchoring strategies, emotional appeals, scarcity cues, and social evidence.

[0014] Optionally, S1 further includes maintaining a mapping strategy matrix between content features and mental account types, and dynamically updating the mapping strategy matrix through a strategy optimization engine; the core of the strategy optimization engine is a collaborative workflow of a context-based multi-armed gambling machine model and an online A / B testing framework, which updates the expected value estimates of each strategy combination in the mapping strategy matrix based on user feedback data through an online learning algorithm.

[0015] Optionally, S1 further includes a context awareness and temporary weight adjustment mechanism, specifically including: monitoring preset key behavioral events as triggers; when the trigger event occurs, generating a temporary mental account weight adjustment vector through a context interpretation model; weighting and normalizing the basic mental account weights based on long-term behavior and the adjustment vector to obtain real-time mental account weights for real-time content matching decisions, wherein the adjustment has time-dependent decay.

[0016] Optionally, in S2, the hierarchical reinforcement learning model under constraints adopts a two-level decision framework, including: an upper-level meta-controller responsible for deciding macro-level changes to the account tree structure; and a lower-level sub-controller responsible for micro-level resource allocation and parameter tuning for each node under a given structure. The training of the model adopts a proximal policy optimization algorithm combined with a constrained policy optimization framework, and integrates resource constraints into the learning process through the Lagrange relaxation method.

[0017] Optionally, S3 specifically includes: maintaining a structured knowledge base of creative elements, in which materials are labeled with appropriate mental account types and historical utility scores;

[0018] Based on the user's real-time mental account type, candidate materials are retrieved from the knowledge base and assembled into a complete promotional content unit by a dynamic content assembler according to the content template corresponding to the account type.

[0019] By integrating user value, contextual value, and content matching degree through a real-time value prediction model, the expected return on display is predicted. Combined with a game theory optimization module and a user fatigue control model, a delivery strategy including bidding and frequency control is generated and executed.

[0020] Optionally, in S4, the integrated anomaly detection model adopts a multi-model parallel processing and decision fusion architecture, specifically running three independent detection channels in parallel: time series prediction and deviation detection channel, unsupervised clustering and outlier detection channel, and deterministic rule filtering channel; the primary anomaly signals generated by the three channels are comprehensively evaluated by a meta-classifier, and the final anomaly judgment conclusion and severity level are output.

[0021] Optionally, after confirming the anomaly, an automated root cause correlation analysis is performed. The analysis constructs a temporary causal graph model containing anomaly nodes, correlation node indicators, and external event factors through a causal discovery algorithm, and uses a Bayesian diagnostic network to calculate the posterior probability of various potential root causes. The posterior probability and the anomaly state are used together as inputs to the contextual multi-armed gambling machine intervention module for decision parameter fine-tuning, strategy switching, structural intervention, or resource reallocation optimization actions.

[0022] Optionally, the method further includes step S5, privacy protection and compliance control:

[0023] Differential privacy noise injection technology is introduced during data collection, model training and data feedback, and the long-term privacy leakage risk is controlled through the privacy budget management module;

[0024] When multiple parties hold different data, federated learning technology is used for joint model training, and homomorphic encryption or secure multi-party computation protocols are used to protect the data privacy of each party during the training process.

[0025] In response to the user's command to disable personalization, immediately disconnect the link between the user profile data and personalized content matching, and switch to a general content strategy.

[0026] Optionally, the method further includes a model knowledge transfer protocol based on the exchange of encrypted model parameters and privacy-preserving summaries, specifically including: the source platform performs importance screening on the pre-trained model parameters and injects noise that meets differential privacy requirements to generate a noisy parameter set and generate an encrypted feature distribution summary;

[0027] The target platform receives the noisy parameter set and the encrypted digest, which are used to initialize the local model and perform domain adaptation.

[0028] When encrypted interactions are allowed, a round of joint fine-tuning for privacy protection based on homomorphic encryption gradients is performed.

[0029] Compared with the prior art, this application has at least the following beneficial effects:

[0030] This application constructs dynamic user profiles through a time-series behavioral encoder and uses machine learning models to categorize users into different mental account types such as "necessities budget" and "emotional reward," transcending the limitations of static tags. Combined with a dynamically updated "content feature-mental account" mapping strategy matrix, it can accurately match users' internal psychological motivations and decision-making patterns, thereby generating or selecting promotional content with stronger emotional resonance and better matching conversion motivation, thus improving the attractiveness of the content and conversion efficiency.

[0031] This application employs a hierarchical reinforcement learning model under constraints, which can automatically explore and generate the optimal multi-level account tree structure based on promotion goals, resource constraints, and user segmentation results. Under the premise of meeting hard constraints such as budget and channels, this model can autonomously make decisions on the creation, merging, splitting, and deletion of the structure and resource allocation through simulation and learning, freeing manual labor from tedious and experience-based structure design, and can dynamically adjust the structure based on real-time performance data to ensure that the system is always in a high-efficiency state.

[0032] This application achieves rapid generation of personalized content based on mental accounting through a structured creative element knowledge base and a dynamic content assembler. By integrating a dynamic bidding and frequency control engine that combines real-time value estimation, game theory optimization, and user fatigue control, it achieves optimal bidding and placement decisions in a two-sided market environment. It organically unifies content creativity, user psychology, market competition, and placement costs, thereby achieving refined strategy execution and maximized returns. Attached Figure Description

[0033] Figure 1 A flowchart illustrating the AI-based automated construction and monitoring method for information promotion account structures provided in this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0035] The AI-based method for automating the construction and monitoring of information promotion account structures provided in this application includes the following steps:

[0036] S1. Multi-dimensional user profile construction and mental account mapping: Collect user behavior data, preference characteristics and historical interaction records, and construct dynamic user profiles through time-series behavior encoders; Based on the user profiles, use machine learning models to determine the mental account type of the user, and establish a three-layer mapping relationship between users, mental account types and promotional content characteristics.

[0037] Specifically, the data sources for dynamic user profiles include, but are not limited to: explicit attribute data (such as device information and basic demographic attributes), implicit behavioral sequence data (such as clickstream, dwell time, search queries, purchase history, and content consumption records), and interactive feedback data (such as likes, favorites, comments, and complaints). These data are cleaned and processed through feature engineering to generate structured user feature vectors.

[0038] Furthermore, a temporal behavior encoder (such as an attention-based recurrent neural network) models user behavior sequences, capturing not only static preferences but also identifying the evolutionary trends and contextual dependencies of their behavioral patterns (e.g., different behavioral patterns on weekdays versus weekends, and during commutes versus at home). Simultaneously, through cross-channel identity recognition and fusion technologies, fragmented user data from different platforms or touchpoints is integrated to form a unified panoramic view. Profile dimensions include, but are not limited to: price sensitivity coefficient, impulsive consumption index, brand loyalty index, content preference vector (e.g., preference for entertainment, knowledge, and promotional content), and real-time contextual status (e.g., geographical location, network environment, and device battery level).

[0039] After obtaining rich user profile features, users are categorized into different mental account types based on a predefined set of mental account determination rules. This set of rules is not fixed but is continuously learned and optimized from historical promotional campaign performance data through machine learning models (such as gradient boosting decision trees or deep neural networks).

[0040] Examples of mental account types include (but are not limited to):

[0041] "Essential Items Budget" account: Users in this account make rational decisions, are sensitive to price, functionality, and quality comparisons, and are risk-averse. Identifying characteristics may include: frequent browsing of product specifications, comparison of different vendors, long purchase decision cycles, and a positive response to promotional terms (such as "essential" or "must-have").

[0042] "Emotional Reward" Accounts: Users view this account as a way to reward themselves, making decisions more emotionally driven, focusing on pleasurable experiences and instant gratification, and being relatively insensitive to price. Identifying characteristics may include: browsing luxury goods or interest categories, active spending at night or on weekends, and high responsiveness to keywords such as "pamper yourself," "limited edition," and "collaboration."

[0043] "Investment Enhancement" accounts: Users focus on long-term returns and self-improvement, prioritizing future value in their decisions. Identifying characteristics may include: consistent interest in education, courses, software tools, and health products; in-depth content reading; and high repurchase rates.

[0044] "Small change transaction" accounts: Users are indifferent to small expenditures, make quick decisions, and are prone to impulsive spending. Identifying characteristics may include: frequent small-amount payments, and rapid response to low-threshold incentives such as "flash sales," "1-yuan purchases," and "red envelopes."

[0045] Calculate the probability distribution of each user's mental account category, rather than a single label, to support hybrid strategies;

[0046] It also maintains a "content feature-mental account" mapping strategy matrix. This matrix defines the most effective combination of features that promotional content should possess for each type of mental account. These features include:

[0047] Information framework: Should the emphasis be on "saving money" (transaction utility) or "gaining experience" (gaining utility)?

[0048] Price anchoring strategy: Does it offer original price comparison, breakdown of average cost per person, or bundled sales schemes?

[0049] Emotional appeal: Rational persuasion or emotional resonance?

[0050] Scarcity indicator: Whether and how to use indicators such as "limited time" and "limited quantity";

[0051] Social evidence: Whether sales figures, positive reviews, etc. are displayed;

[0052] When it's necessary to push content to a specific user, the system first calculates the user's current dominant mental account type (or probability distribution) based on their dynamic profile. Then, it queries the mapping strategy matrix to generate content creation instructions or intelligently matches the most suitable content materials from the content library. For example, for users identified as "emotional reward" accounts, the system automatically selects visual materials and copy that emphasize unique experiences, aesthetic design, and pleasure, while avoiding overemphasizing cost-effectiveness.

[0053] Furthermore, the dynamic update of the mapping strategy matrix is ​​performed by an independent strategy optimization engine, the core of which is a collaborative workflow of a context-based multi-armed gambling machine model and an online A / B testing framework.

[0054] At each content promotion decision point, a state context vector is constructed based on the user's current profile and the calculated dominant mental account type. Upon receiving this vector, the strategy optimization engine does not directly adopt the historically best strategy. Instead, it selects a strategy from multiple candidate content strategies (i.e., "arms") corresponding to that account type in the strategy matrix with a certain exploration probability. The exploration probability is dynamically calculated using Thompson Sampling or Upper Confidence Bound (UCB) algorithms, aiming to balance "utilizing" known effective strategies with "exploring" potentially better strategies.

[0055] Subsequently, content versions employing different strategies are delivered to user groups with similar characteristics in a scientifically distributed manner using an integrated A / B testing framework. This framework rigorously records the complete context of each display instance, the strategies employed, and subsequent multi-dimensional user feedback signals. These feedback signals include not only immediate interaction metrics (such as click-through rate), but also conversion metrics aligned with business goals (such as order rate and payment amount), and long-term value metrics processed with delayed feedback (such as changes in user lifetime value and retention rate).

[0056] These feedback data streams are input in real time into the reward function calculation module of the strategy optimization engine. This module combines the feedback from multiple indicators into a scalarized, instantaneous reward value based on preset business objective weights. The contextual multi-armed gambling machine model then uses this reward value and corresponding contextual information to update its internal parameters through online learning algorithms (such as LinUCB or neural network-driven gradient descent), thereby dynamically adjusting the expected value estimates of each "account-content" strategy combination in the strategy matrix.

[0057] This process forms a complete "decision-feedback-learning" closed loop. The strategy matrix can be updated incrementally, that is, the parameters are fine-tuned after each batch of feedback data is received; or it can be configured for periodic retraining, that is, the model is retrained periodically using accumulated historical data to discover deeper patterns. Through this mechanism, the mapping strategy matrix can continuously adapt to changes in market trends, user preference shifts, and changes in the competitive environment, achieving autonomous online evolution and optimization of promotion strategies.

[0058] Furthermore, to achieve dynamic and situational adaptability in mental accounting assessment, a situational awareness and temporary weight adjustment mechanism is added. Independent of the basic user profile update process, this mechanism focuses on capturing and analyzing the user's real-time behavioral event flow, and accordingly makes short-term, targeted adjustments to the user's current mental accounting weight distribution. Specifically, the process includes the following:

[0059] A series of key behavioral events with high contextual information content are preset as triggers. These events include not only explicit transaction behaviors (such as "completing a large payment" or "renewing a subscription service"), but also specific browsing behavior sequences (such as "leaving after browsing multiple high-end product detail pages" or "making multiple queries in a price comparison tool"), as well as external contextual signals (such as accessing a Wi-Fi network, geographical location entering a commercial area, or the device detecting night mode). Each trigger is associated with a contextual strength coefficient and an impact time window.

[0060] Upon detecting a triggering event, this module immediately extracts the event's contextual features to form a contextual signal vector. For example, for a "transaction completed" event, the vector includes the transaction amount, product category, payment method, and premium percentage relative to the user's historical average transaction amount. This vector is then input into a lightweight contextual interpretation model. This model can be a rule-based classifier or a small neural network, and its function is to determine the direction of the user's psychological state shift that this contextual signal may imply. For example, the model might determine, based on rules, that an event like "a 'home furnishing and building materials' expenditure far exceeding the daily average" likely means that the user has just completed a rational, high-investment decision regarding a "necessity" or "family investment."

[0061] Based on the output of the context interpretation model, a temporary mental accounting weight adjustment vector is generated. This vector indicates the incremental adjustment to be made to the probability distribution of the user's basic mental accounts within the current time window. For example, for the aforementioned large-amount purchase of necessities, the adjustment vector might be: {"Necessities Budget Account": -0.2, "Emotional Reward Account": +0.3, "Investment Value-Added Account": 0, "Small Change Transaction Account": -0.1}, indicating a temporary reduction in the weight of the "Necessities Account" and a significant increase in the weight of the "Emotional Reward Account";

[0062] The adjustment algorithm employs weighted superposition and normalization: the basic weights and adjustment vectors are combined in a preset ratio (e.g., 7:3), and then re-normalized into a probability distribution. This adjustment is time-sensitive, with a decay period set for each adjustment. Its influence decays exponentially over time until it reaches zero, and the user's mental account profile will eventually return to the basic state determined by long-term behavior.

[0063] When making content matching decisions, the content recommendation engine prioritizes using real-time mental account weights adjusted by this module, which can capture users' instantaneous and contextualized psychological states. For example, in the case above, the user may no longer be exposed to cost-effectiveness ads shortly after the transaction, but rather to promotional content for high-end dining, short trips, or affordable luxury goods, accurately matching their potential "self-compensation" psychological motivation. This mechanism greatly enhances the emotional resonance and timing effectiveness of promotional content at critical decision-making moments;

[0064] S2. Automated hierarchical design of promotion account structure: Based on the promotion goals, resource constraints and the user mental account hierarchical results obtained in step S1, a hierarchical reinforcement learning model under constraints is used to automatically generate a multi-level promotion account structure, in which each account node encapsulates specific content strategies, delivery channels, resource budgets and target audience rules.

[0065] The multi-tiered promotion account structure is a tree-like or graph-like topology. Each account node is an independent strategy execution unit, encapsulating complete promotional elements, specifically including:

[0066] Strategy Configuration: Points to the specific strategy ID in the "User-Mental Account-Content" mapping strategy matrix, which determines the content creation and matching logic;

[0067] Delivery channel configuration: The specific advertising platform or content distribution channel to be bound (such as feed ads, search engine marketing, social media push, SMS channel) and its corresponding technical parameters (such as channel API interface identifier);

[0068] Content format template: Specifies the basic format (such as text, short video, interactive H5) and creative element constraints for the content to be pushed;

[0069] Resource budget and bidding strategy: the budget amount allocated to the node, the consumption rate control rules, and the bidding function (such as CPC, oCPM bidding logic) in the bidding environment.

[0070] Target Audience Filter: Audience targeting rules based on user profiles and real-time mental account weights, used to filter the subset of users that the node is responsible for reaching from the user pool;

[0071] The hierarchical relationship in the structure represents the generalization and refinement of the strategy. High-level nodes (such as root nodes or trunk nodes) usually correspond to a broad user group and general brand strategies, with a larger resource pool; low-level nodes (leaf nodes or branch nodes) correspond to segmented mental accounts or precise strategies in specific situations, with more targeted strategies, which are easier to track and iterate quickly.

[0072] The goal of the constrained hierarchical reinforcement learning model is to autonomously explore and generate the optimal promotion account topology while satisfying multi-dimensional resource constraints. The model employs a two-tier decision-making framework: the upper layer is the meta-controller, responsible for determining macro-level changes to the account tree structure; the lower layer is the sub-controller, responsible for micro-level resource allocation and parameter tuning for each node given the structure.

[0073] The model's state space is defined as a complete description of the current account structure, including the attribute vectors of all nodes (strategy ID, channel binding, budget balance, historical performance indicators), the topological relationships between nodes, and external input information (total target, total budget, list of available channels, real-time user stratification statistics). Its action space consists of a set of discrete and continuous actions. Discrete actions include: creating a new child node under a specified parent node, merging two specified nodes, splitting a node into multiple child nodes, and deleting a specified node. Continuous actions include: adjusting the budget allocation weights between any nodes and modifying the bidding strategy parameters of a specific node.

[0074] To evaluate the effectiveness of the actions, the algorithm relies on an offline simulator built on historical data. This simulator can simulate the estimated exposure, user interaction, and conversion chain of each strategy on the corresponding channel, given an account structure and resource input. The model executes actions in the simulator and receives reward values ​​calculated by a composite reward function. This function integrates multiple objectives, such as total conversions, customer acquisition cost, and user satisfaction metrics (e.g., reducing negative feedback from ad harassment), and linearly weights them using preset weights. Simultaneously, the function includes a penalty term that imposes a significant negative reward for any violation of hard constraints (e.g., budget overruns, channel quota exceeding limits).

[0075] The handling of constraints is integrated into the learning process using Lagrange relaxation. Specifically, resource constraints (such as total budget and daily consumption limit per channel) are transformed into penalty terms, and learnable Lagrange multipliers are introduced. The model's objective becomes maximizing the original reward while minimizing the penalty for constraint violation. The multipliers themselves are also updated using gradient methods, thereby dynamically adjusting the emphasis on constraints during training.

[0076] The model is trained using a Proximal Policy Optimization (PPO) algorithm combined with a Constrained Policy Optimization (CPO) framework. Through extensive trial and error with the simulator, the model learns which structural changes and resource adjustments should be made under what conditions to maximize cumulative rewards in the long run while ensuring constraints are met. The trained model then possesses the ability to automatically generate and dynamically adjust a high-performance, compliant hierarchical structure of promotional accounts based on real-time input.

[0077] At startup, it can begin with a simple structure (such as a single node) or be initialized based on a template of successful historical cases. During the promotion campaign, the engine periodically (or triggered by key events) re-runs the structure generation algorithm, dynamically optimizing the existing structure based on the latest user mental account distribution data and real-time performance data of each node (such as consumption rate, conversion cost, and user feedback).

[0078] Optimization operations include:

[0079] Performance-driven merging: When the target audience characteristics of two leaf nodes converge and their strategies are similar, but their respective performance has not reached the scale effect threshold, they are automatically merged into one node for centralized budgeting;

[0080] Opportunity-driven splitting: When a subset of the audience within a node performs significantly better or worse than the whole, it is automatically split into independent child nodes in order to implement a differentiation strategy;

[0081] Weight rebalancing: Based on the real-time return on investment (ROI) of each node, dynamically adjust its budget allocation weight, and tilt resources toward high-efficiency nodes;

[0082] The resulting account structure is deployed to the execution systems of various promotion channels through a standardized configuration interface, enabling automated distribution and delivery of strategies.

[0083] S3, Intelligent Content Matching and Execution of Delivery Strategy: Based on the real-time mental account type of users obtained in S1 and the strategy defined by the account nodes in S2, it uses natural language processing and image recognition technology to match or generate personalized promotional content, and dynamically optimizes the delivery strategy in combination with the two-sided market model, and executes delivery through the channels bound to the account nodes in S2.

[0084] Specifically, a structured knowledge base of creative elements needs to be maintained. This base stores a large amount of deconstructed content materials, including text snippets, visual materials (images, video clips), interactive components, etc. Each material is labeled with rich metadata, a key element of which is the corresponding mental accounting type and its historical utility score under that type;

[0085] When content needs to be generated for a specific promotional request, the engine first retrieves a batch of highly relevant candidate materials from the knowledge base based on the user's mental accounting results. Then, a dynamic content assembler combines these materials into complete promotional content units (such as in-feed ad images and text, short video ad scripts) according to preset content templates corresponding to the account type. The assembly process is not simply piling things up; it uses natural language generation technology to smoothly integrate the text and utilizes image recognition and layout algorithms to ensure the harmony and prominence of visual elements. For example, for an "emotional reward" account, the template might specify large images to highlight the atmosphere and text emphasizing unique experiences and instant gratification; while for a "necessities budget" account, the template might require a clear display of parameter comparisons, price advantages, and guarantee information.

[0086] Content delivery strategies (including bidding, frequency, timing, and channel priority) are generated in real time by an independent dynamic bidding and frequency control engine. This engine treats each promotion opportunity as a micro-level two-sided market: the demand side is the promotion objective, and the supply side is the immediate ad space inventory or user attention;

[0087] At the core of the engine is a real-time value prediction model that integrates user value (based on their mental account type, historical conversion rate, and lifetime value prediction), contextual value (such as current time period and geographic location), and content relevance to predict the expected return on this impression. Based on this, the engine runs a game theory optimization module that simulates the competitive environment with other advertisers on the platform (or other competing nodes under this account). By solving a simplified non-cooperative game equilibrium (such as a Bayesian Nash equilibrium approximation under a generalized second-price auction), this module outputs an optimal bidding function under the current market environment. This function considers not only the direct value of this click or conversion but also the long-term competitive landscape and budget smoothing.

[0088] Simultaneously, a user fatigue and frequency control model works in parallel. This model tracks the frequency with which each user is exposed to specific ads or similar content and the recent interaction decay, dynamically calculates the current optimal contact interval and daily exposure limit to ensure that the user experience is not excessively disturbed, and incorporates this as a hard constraint or penalty into strategy generation.

[0089] The generated campaign strategies (including specific bids and target frequencies) and the assembled content are packaged into standard instructions and sent to the execution ends of various promotion channels. After strategy execution, end-to-end data such as exposure, clicks, and conversions are collected in real time as feedback signals. These signals are used for two types of updates: 1) Content element utility updates: updating the utility scores of relevant materials in the creative element knowledge base under the corresponding mental accounts; 2) Strategy model updates: serving as training data to drive the online learning of the real-time value prediction model and game theory optimization module, enabling them to adapt to dynamic changes in the market environment.

[0090] S4. Real-time monitoring and feedback-driven structure optimization: Real-time collection of performance indicator streams of each node in the account structure generated in step S2, identification of anomalies through an integrated anomaly detection model, and generation of optimization actions in combination with the contextual multi-armed gambling machine model to dynamically adjust the account structure, resource allocation or content strategy.

[0091] Specifically, to further achieve the continuous optimal operation of the promotion account structure, an independent real-time monitoring and optimization closed-loop mechanism is also needed. This mechanism collects performance indicator stream data of each node in the account structure across multiple dimensions at high frequency (e.g., every minute), and drives the dynamic adjustment of the structure through an automated "monitoring-diagnosis-optimization" process. The specific process includes the following:

[0092] Data is pulled in real-time from the application programming interfaces (APIs) of various promotional channels and then standardized and aggregated. A multi-dimensional time-series metric set is built for each account node, with core metrics including but not limited to: impressions, click-through rate (CTR), conversion rate, average cost per conversion (CPC), burn rate, and user interaction feedback ratings (such as likes, dislikes, and reasons for closing). These raw metric streams pass through a feature extraction layer to calculate insightful derived features, such as: the deviation of the metric from its own historical baseline (such as the average of the same period in the past 7 days), changes in the cooperative relationship between different metrics (such as an increase in burn rate but a simultaneous decrease in CTR), and composite metrics that conform to the preset business logic (such as "value-for-cost ratio," i.e., the expected conversion value brought by a unit of consumption).

[0093] The characterized multidimensional indicator time series data is fed into an integrated anomaly detection and diagnosis engine in real time. This engine employs a multi-model parallel processing and decision fusion architecture to achieve highly robust anomaly detection with a low false alarm rate.

[0094] The integrated detection module operates three independent detection channels in parallel: the first channel is the time-series prediction and deviation detection channel. This channel maintains a dedicated long short-term memory network prediction model for each key performance indicator (KPI). This model takes the historical sequence of the KPI (typically containing periodic and trend characteristics) as input, outputs the predicted value and prediction confidence interval for the next time point in real time, calculates the standardized residual between the current actual value and the predicted value, and compares it with a threshold dynamically calculated based on the historical residual distribution. When the residual exceeds the threshold for multiple consecutive periods, or when the magnitude of a single exceedance reaches an extreme level, a primary anomaly signal is generated, along with a prediction deviation score.

[0095] The second channel is an unsupervised clustering and outlier detection channel. This channel combines the multi-dimensional derived features (such as consumption growth rate, interaction rate, and cost ratio) of all account nodes in the same time slice into a high-dimensional feature vector, and uses an online clustering algorithm (such as incremental K-means or a density-based DBSCAN variant) for real-time segmentation. Node feature vectors located in sparse regions or whose distance from their respective cluster centers exceeds a threshold are identified as spatial outliers, generating a primary anomalous signal in another dimension.

[0096] The third channel is a deterministic rule filtering channel. This channel loads a series of hard and soft rules explicitly defined by business logic. Hard rules (such as "the consumption of any node in a single minute shall not exceed 10% of its daily budget") generate a high-priority anomaly signal once triggered. Soft rules (such as "the click-through rate has decreased by more than 30% compared to the same period of the previous day") generate medium-to-low-priority signals when triggered, and require cross-validation with results from other channels.

[0097] The primary anomaly signals, scores, and metadata (such as the triggered rule ID and the deviation dimension) generated by the three channels mentioned above are fed into a lightweight meta-classifier. This meta-classifier is typically a gradient boosting tree model, trained on a large amount of historical labeled data (normal vs. anomalous events). It is responsible for comprehensively evaluating the evidence from different channels and outputting a final anomaly determination, a comprehensive severity level (such as "warning," "severe," or "fatal"), and the main representation dimension of the anomaly involved.

[0098] Once an anomaly is confirmed by the meta-classifier, the automated root cause correlation analysis process is immediately activated. This process first defines the analysis scope, typically including the anomaly node itself, its direct parent-child and sibling nodes in the account structure (based on the assumption that structural correlation may suggest common causes), and other nodes whose indicator performance fluctuates significantly within the same time period. The core of the analysis is to apply causal discovery and probabilistic inference techniques to construct a temporary causal graph model containing key indicator variables and potential external dependent variables. Using constraint-based causal discovery algorithms (such as the PC algorithm or FCI algorithm), the conditional independence relationships between node indicators within the above scope are analyzed, and possible causal arrow directions are preliminarily inferred. Simultaneously, an external event knowledge base is queried. This knowledge base continuously records known events that may affect promotional effectiveness, such as announcements of interface failures in advertising channels, algorithm strategy updates, the launch of promotional activities by major competitors, and even major social news events. These external events are incorporated into the causal graph as candidate confounding factors or direct causes.

[0099] Finally, by using a pre-built Bayesian diagnostic network or performing graph-based probabilistic reasoning, the posterior probabilities of various potential root causes (e.g., "general decline in traffic quality of specific advertising channels", "failure of content strategy templates for a certain mental account", "intensified market competition leading to a deterioration of the bidding environment") are calculated. The probabilistic diagnostic results, together with the current state of the abnormal nodes, constitute the input for subsequent optimization decisions, providing them with precise action directions.

[0100] Instead of directly executing preset rules based on the diagnostic results, a context-based multi-armed gambling machine intervention module makes decisions on optimized actions. This module uses the current "system state" (including anomaly type, root cause probability distribution, current account structure, and remaining budget) as context. Its selectable "arms" (i.e., actions) are a rich set of actions covering different granularities:

[0101] Parameter fine-tuning arm: Adjusts the bidding coefficient and budget consumption rate limit for abnormal nodes;

[0102] Policy switching arm: Switches the content policy bound to a node to the alternative policy in the mapping policy matrix;

[0103] Structural intervention arm: Triggers the account structure generation engine to perform lightweight reorganization of local structures (such as splitting the node or merging it with another node).

[0104] Resource reallocation arm: Transfers budget weights between nodes;

[0105] Based on the historical rewards of each arm (i.e. the degree of improvement of key metrics in subsequent time windows after the action is performed), the module uses the Thompson sampling algorithm to select the action with the highest expected return in the current context and reserves a portion of the throughput to explore other potentially effective actions.

[0106] The effectiveness of all optimization actions is continuously tracked and fed back as new experiential data to the aforementioned modules. The anomaly detection model is incrementally updated based on the new data distribution; the experience of the root cause association model is accumulated; and the reward estimates for each arm of the multi-armed gambling machine are updated in real time. This forms a complete autonomous closed loop from perception to decision-making to learning, ensuring that the account structure and strategies can adapt to rapidly changing internal and external environments.

[0107] S5. Privacy Protection and Compliance Control: Differential privacy noise injection and federated learning techniques are introduced during the data collection in step S1, model training in step S3, and data feedback in step S4. At the same time, in response to the user's "one-click to turn off personalization" command, the link of the user profile data used for personalized content matching is immediately cut off, and the general content strategy is switched.

[0108] In the initial stages of data collection and user profile construction, the system introduces centralized or localized differential privacy techniques. For aggregated statistical data gathered on the server (such as the average click rate of a certain mental account group), calibrated Laplace noise or Gaussian noise is injected before publication to ensure that the presence or absence of individual user data does not significantly affect the statistical results, thus satisfying a strict mathematical definition of privacy. For scenarios requiring preliminary calculations on the client side (such as calculating the local embedding vector of user behavioral features), a local differential privacy model is adopted. Users perturb the data on their devices before uploading, and the server only receives the ambiguous information with added noise. The system tracks and allocates the privacy budget consumed by each user or each query through a privacy budget management module. This ensures that the long-term accumulated risk of privacy breaches is controllable;

[0109] When participating parties (defined as "clients," such as data partitions of different business units or regional servers that meet data sovereignty requirements) hold datasets with the same feature space but different user groups, a horizontal federated learning process is initiated. In each training cycle, the central coordinating server distributes the current global model parameters to a selected subset of clients. Each selected client uses its local user data to independently perform forward and backward propagation computations, generating a local model update (usually gradients or parameter differences) for the global model. To protect the information contained in the update, the client encrypts its update using additive homomorphic encryption or a secure multi-party computation protocol before sending it. The encrypted local update is then uploaded to the central server.

[0110] The central server runs a secure aggregation protocol that aggregates updates to the ciphertext state directly without decrypting individual client updates, resulting in an encrypted global model update. Subsequently, a pre-defined decryption process (potentially involving multi-key or threshold decryption) yields the aggregated plaintext global model update, which is then applied to the global model to complete the current iteration. This process ensures that the central server can never obtain plaintext gradient information from any single client, thus preventing it from retrieving the original data.

[0111] When the data held by the participants involves different feature dimensions of the same user group (for example, party A has user browsing and click behavior data, while party B has user transaction and average order value data), the system adopts vertical federated learning;

[0112] First, both parties use a technique based on the intersection of privacy-preserving sets to confirm shared user identifiers in an encrypted state without revealing their respective unique user lists. For shared user samples, both parties need to collaborate on model training. Taking the training of a neural network model as an example, the network is vertically partitioned: the lower-level parts related to their respective features are computed locally by the data holder, generating intermediate results. These intermediate results need to be exchanged through an encrypted channel to calculate the loss and propagate gradients;

[0113] During this process, all intermediate data involved in cross-party transmission is processed using homomorphic encryption or obfuscation circuits to ensure that the receiver cannot parse the original feature information of the other party. Gradient calculation is completed through a secure multi-party computation protocol, enabling each party to obtain the gradient updates required for its own model portion without knowing the feature data or gradient details held by the other party. Finally, each party securely updates its locally held model parameters using its decrypted gradients, collectively forming a complete and functionally consistent joint model, while the original data of any party remains within its local closed loop.

[0114] To improve the efficiency and robustness of federated learning, the following mechanisms will be introduced:

[0115] Client selection and incentives: In each round of training, clients are dynamically selected based on their computing resources, data quality, historical contributions, and network conditions, and incentives are given for effective contributions;

[0116] Enhanced differential privacy: When updating locally on the client side, additional noise that meets differential privacy requirements can be injected to provide double protection for privacy and resist more powerful attacks;

[0117] Anomaly and malicious client detection: The central server uses statistical methods (such as the norm distribution of the update vector and cosine similarity with other clients) to detect clients that may provide low-quality or malicious updates and excludes them from aggregation, thus ensuring the security and convergence of the global model.

[0118] Model version management and consistency coordination: The system maintains the version history of the global model and handles the problem of missing updates due to network latency or client disconnection, ensuring the consistency of the final model state;

[0119] Through the implementation of the above-mentioned federated learning mechanism, the system can make full use of the value of data distributed in various places within a legal and compliant framework, continuously optimize key algorithm models, and at the same time fundamentally eliminate the privacy leakage risks associated with cross-border and cross-system transmission and centralized storage of original user data.

[0120] The user interface provides a clear privacy control panel, through which users can exercise their right to "opt out" or "withdraw consent." When a user triggers the "one-click to disable personalization" command, the system immediately executes a series of linked operations:

[0121] Identifier processing: Remove the user's unique identifier (such as Cookie ID, Device ID) from the active target audience pool used for personalized recommendations, or associate it with an anonymous group that only accepts non-personalized content pushes;

[0122] Data processing pipeline switchover: Immediately cease any collection of tracking data for updating the user's personalized profile. Subsequent behavioral data, if necessary for the service, will only be used for overall service optimization in a de-identified and aggregated manner.

[0123] Forced content strategy downgrade: When the content recommendation engine receives a user's request, it will ignore the user's detailed profile and mental account classification, and instead match the content in the general content strategy pool based on the simplest context such as the user's location and device type, to ensure that the pushed content is completely independent of the user's personal data.

[0124] Data deletion and retention compliance: Based on user selection and applicable legal requirements (such as GDPR, CCPA), automatically trigger the deletion or anonymization process of related data and fully record the event in the audit log;

[0125] The compliance rules engine encodes the specific requirements of relevant privacy regulations in the target market (such as the scope of user consent, data minimization principles, and storage periods). This engine monitors and scans data collection points, data processing logic, model training operations, and content recommendation decisions in real time, issuing interceptions or alerts for any potential violations. Simultaneously, it automatically generates detailed records of data processing activities, providing verifiable compliance evidence for possible regulatory inquiries.

[0126] Furthermore, in scenarios where the source platform and target platform are subject to strict data isolation policies or privacy regulations, making direct interaction of any form of raw data or even encrypted samples impossible, a migration protocol based on the exchange of encrypted model parameters and privacy-preserving digests is designed and implemented to achieve secure cross-platform migration of model knowledge.

[0127] Specifically as follows:

[0128] On the source platform side: Secure encapsulation and desensitized summary generation of model knowledge. The source platform first performs necessary knowledge distillation and structured desensitization processing on the pre-trained model to be transferred (e.g., the mental accounting judgment model). Specific steps include:

[0129] Parameter Importance Screening and Noise Injection: Model pruning and parameter importance assessment techniques (such as sensitivity analysis based on the Hessian matrix) are employed to screen out a subset of core parameters that significantly contribute to model performance. For this subset of parameters, Gaussian or Laplace noise that meets differential privacy requirements is injected according to a pre-defined privacy budget (ε, δ), generating a noisy parameter set. This process ensures that even if the parameters themselves are leaked, it is impossible to effectively reverse engineer the specific user data of the source platform.

[0130] Generating a privacy-preserving feature distribution summary: To assist the target platform in effective domain adaptation, the source platform needs to provide a statistical summary of the source domain data features. This summary is not the original feature values, but is achieved by performing encrypted random projection on the feature space or generating an aggregated statistical histogram based on k-anonymization and differential privacy. For example, user feature vectors are mapped to a low-dimensional space through a randomly generated projection matrix that can be securely shared subsequently, and then the mean, covariance, and other statistical measures with added noise are calculated for the data distribution in this low-dimensional space to form an encrypted summary;

[0131] The processed noisy parameter set and encrypted digest are transmitted to the target platform via a secure channel. The transmission process can be recorded in a blockchain or security log for auditing purposes, ensuring the immutability and traceability of the migration event.

[0132] After receiving the migration materials, the target platform will execute the following process:

[0133] Parameter integrity verification: Using a pre-agreed digital signature or hash verification mechanism, the integrity and authenticity of the received parameters and digest are verified to prevent tampering during transmission;

[0134] Model Reconstruction and Initialization: Using the received noisy parameter set, initialize the corresponding model architecture for the target platform. Since the parameters have been injected with noise, this model can be viewed as a starting point that has been pre-trained on the "fuzzy" source domain;

[0135] Domain Alignment Initialization Based on Encrypted Digest: Using the received encrypted feature distribution digest, the target platform can initialize its domain adaptation module. For example, the digest can be used to calculate a difference estimate between the preliminary feature distributions of the source and target domains, and the weights of the domain discriminator can be initialized accordingly, or an initial scaling factor for the maximum mean difference loss can be set, thereby guiding the subsequent fine-tuning process to align the distributions more efficiently.

[0136] If very limited encrypted interaction is permitted, a round of privacy-preserving joint fine-tuning can be performed. The target platform calculates the model gradient locally using its data, encrypts the gradient using homomorphic encryption, and sends it to the source platform. The source platform aggregates the gradient in the encrypted state (possibly combined with its own limited secure computation) and sends the encrypted aggregated gradient back. The target platform decrypts the gradient and updates its model. This process is entirely performed in encrypted form; neither party knows the other's specific gradient values, achieving closer collaborative training without data leakage.

[0137] Using the above methods, the system achieves effective and secure migration of core strategy model knowledge by transmitting only model parameters and statistical summaries that have undergone strict privacy processing, while maintaining absolute isolation between the original data of the source platform and the target platform. This provides a feasible technical path for compliance technology collaboration across business units, regions, or partner companies.

[0138] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. An AI-based method for automated construction and monitoring of information promotion account structures, characterized in that, The method includes: S1. Multi-dimensional user profile construction and mental account mapping: Collect user behavior data, preference characteristics and historical interaction records, and construct dynamic user profiles through time-series behavior encoders; Based on the dynamic user profiles, use machine learning models to determine the mental account type of the user, and establish a three-layer mapping relationship between users, mental account types and promotional content characteristics. S2. Automated hierarchical design of promotion account structure: Based on the promotion goals, resource constraints and the user mental account hierarchical results obtained in S1, a hierarchical reinforcement learning model under constraints is used to automatically generate a multi-level promotion account structure, in which each account node encapsulates specific content strategies, delivery channels, resource budgets and target audience selection rules based on user profiles and mental account weights. S3. Intelligent content matching and delivery strategy execution: Based on the user's real-time mental account type obtained in S1 and the strategy defined by the account node in S2, natural language processing and image recognition technology are used to match or generate personalized promotional content, and the delivery strategy is dynamically optimized in combination with a two-sided market model, and the delivery is executed through the channels bound to the account node in S2. S4. Real-time monitoring and feedback-driven structural optimization: Real-time collection of performance indicator streams of each node in the account structure generated in S2, identification of anomalies through an integrated anomaly detection model, and generation of optimization actions in combination with the contextual multi-armed gambling machine model to dynamically adjust the account structure, resource allocation or content strategy.

2. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 1, characterized in that, In S1, the temporal behavior encoder is a recurrent neural network based on an attention mechanism, used to model user behavior sequences to capture the evolution trend and context dependence of behavior patterns; the three-layer mapping relationship is specifically a mapping from user to mental account type, and a mapping from mental account type to promotional content features, wherein the promotional content features include information frames, price anchoring strategies, emotional appeals, scarcity cues, and social evidence.

3. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 2, characterized in that, S1 further includes maintaining a mapping strategy matrix between content features and mental account types, and dynamically updating the mapping strategy matrix through a strategy optimization engine; the core of the strategy optimization engine is a collaborative workflow of a context-based multi-armed gambling machine model and an online A / B testing framework, which updates the expected value estimates of each strategy combination in the mapping strategy matrix based on user feedback data through an online learning algorithm.

4. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 1, characterized in that, The S1 also includes a context awareness and temporary weight adjustment mechanism, specifically including: monitoring preset key behavioral events as triggers; when the trigger event occurs, generating a temporary mental account weight adjustment vector through a context interpretation model; weighting and normalizing the basic mental account weights based on long-term behavior and the adjustment vector to obtain real-time mental account weights for real-time content matching decisions, which have time-dependent decay.

5. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 1, characterized in that, In S2, the hierarchical reinforcement learning model under the constraints adopts a two-level decision framework, including: an upper-level meta-controller responsible for deciding the macro-level changes to the account tree structure; and a lower-level sub-controller responsible for micro-level resource allocation and parameter tuning for each node under the given structure. The training of the model adopts a proximal policy optimization algorithm combined with a constraint policy optimization framework, and integrates resource constraints into the learning process through the Lagrange relaxation method.

6. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 1, characterized in that, Specifically, S3 includes: maintaining a structured knowledge base of creative elements, in which materials are labeled with appropriate mental account types and historical utility scores; Based on the user's real-time mental account type, candidate materials are retrieved from the knowledge base and assembled into a complete promotional content unit by a dynamic content assembler according to the content template corresponding to the account type. By integrating user value, contextual value, and content matching degree through a real-time value prediction model, the expected return on display is predicted. Combined with a game theory optimization module and a user fatigue control model, a delivery strategy including bidding and frequency control is generated and executed.

7. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 1, characterized in that, In S4, the integrated anomaly detection model adopts a multi-model parallel processing and decision fusion architecture, specifically running three independent detection channels in parallel: time series prediction and deviation detection channel, unsupervised clustering and outlier detection channel, and deterministic rule filtering channel; the primary anomaly signals generated by the three channels are comprehensively evaluated by a meta-classifier, and the final anomaly judgment conclusion and severity level are output.

8. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 7, characterized in that, After confirming the anomaly, an automated root cause correlation analysis is performed. The analysis constructs a temporary causal graph model containing anomaly nodes, correlation node indicators, and external event factors through a causal discovery algorithm, and uses a Bayesian diagnostic network to calculate the posterior probability of various potential root causes. The posterior probability and the anomaly state are used together as inputs to the contextual multi-armed gambling machine intervention module for decision parameter fine-tuning, strategy switching, structural intervention, or resource reallocation optimization actions.

9. The method for automated construction and monitoring of information promotion account structure based on AI according to any one of claims 1-8, characterized in that, The method also includes step S5, privacy protection and compliance control: Differential privacy noise injection technology is introduced during data collection, model training and data feedback, and the long-term privacy leakage risk is controlled through the privacy budget management module; When multiple parties hold different data, federated learning technology is used for joint model training, and homomorphic encryption or secure multi-party computation protocols are used to protect the data privacy of each party during the training process. In response to the user's command to disable personalization, immediately disconnect the link between the user profile data and personalized content matching, and switch to a general content strategy.

10. The method for automated construction and monitoring of information promotion account structure based on AI according to claim 9, characterized in that, The method also includes a model knowledge transfer protocol based on the exchange of encrypted model parameters and privacy-preserving summaries, specifically including: the source platform performs importance screening on the pre-trained model parameters and injects noise that meets differential privacy requirements to generate a noisy parameter set and generate an encrypted feature distribution summary; The target platform receives the noisy parameter set and the encrypted digest, which are used to initialize the local model and perform domain adaptation. When encrypted interactions are allowed, a round of joint fine-tuning for privacy protection based on homomorphic encryption gradients is performed.

Images