Personalized pushing and intelligent distribution system and method

Abstract

The invention discloses a personalized push and intelligent distribution system and method. The system comprises a data acquisition module, a content module, an intelligent decision module, a distribution execution module and a monitoring optimization module. Wherein the data acquisition module sends a feature vector to the intelligent decision-making module; the content module sends content metadata and an A / B experiment label to the intelligent decision-making module; the intelligent decision-making module sends the recommendation content, the distribution strategy and the A / B experiment label to a distribution execution module; the distribution execution module displays the received data to the user, collects feedback information of the user, and sends the feedback information of the user to the monitoring optimization module; and the monitoring optimization module analyzes feedback information of the user and sends an analysis result to the intelligent decision-making module, so that circulation is formed until the user completes the task of the round. According to the method, accurate distribution and personalized pushing are realized through multi-dimensional attribute and fine granularity decomposition.

Description

TECHNICAL FIELD

[0001] The present application relates to a system for personalized pushing and intelligent distribution, and also relates to a corresponding method, belonging to the technical field of digital medicine. BACKGROUND

[0002] In the fields of digital medicine, education, and elderly care, the accuracy, flexibility, and scalability of intelligent distribution systems are increasingly prominent. In the context of digital medicine, for ADHD children, elderly patients with cognitive impairment, and people with anxiety and depression, it is necessary to integrate training tasks, patient education videos, age-appropriate interfaces, and physiological indicator monitoring to achieve dynamic intervention based on real-time biological feedback and contextual data. However, existing intelligent distribution systems are often limited to a single disease or content type, such as supporting only ADHD training tasks or psychological health video recommendations, and the content is not annotated with multiple dimensions, making it difficult to break down into reusable units, resulting in the inability to flexibly combine content based on user preferences and real-time state. In the field of education, adaptive learning requires systems to recommend personalized tasks based on students' knowledge levels, learning styles, and real-time attention data, but existing platforms often rely on static historical performance and lack the ability to dynamically perceive learning state and adjust task difficulty in real time. Elderly health management requires the integration of physiological indicators and operation habits to push health monitoring tasks, nutrition videos, and emergency notification functions, while traditional systems often use fixed pushing strategies and cannot dynamically adjust exercise recommendations based on blood pressure fluctuations or optimize interface layouts based on declining vision.

[0003] General recommendation systems can achieve high-concurrency content distribution, but they rely on shallow learning patterns based on user behavior data, lack professional logic in medical and educational verticals, and cannot meet medical data security compliance requirements. A / B testing tools face granularity issues in medical scenarios, as they can only test complete pages or tasks and cannot support combination experiments with fine-grained units. Moreover, causal inference delays are usually in seconds, making it difficult to meet the real-time needs of medical scenarios. In addition, existing architectures often lack distributed computing capabilities when dealing with millions of concurrent users, leading to increased response delays and inadequate adaptation to emerging technology scenarios, making it difficult to support multi-modal input and low-latency inference.

[0004] Overall, existing intelligent distribution systems have significant limitations in supporting multiple content, controlling multiple dimensions, distributing fine-grained units, real-time dynamic optimization, high-concurrency performance, data security, and cross-scenario expansion capabilities, making it difficult to meet the increasingly complex personalized needs of multiple fields. Therefore, technological breakthroughs are needed to build more flexible, accurate, and secure intelligent distribution architectures.

[0005] In the Chinese invention patent with the authorization announcement number CN118467851B, a kind of artificial intelligence data search and distribution method and system are disclosed, the scheme is by obtaining the search request input by user, the semantic extension is carried out to search request, obtains target extension search keyword, according to target extension search keyword and the multi-modal semantic index of pre-construction, semantic search is carried out, obtains preliminary search result;Through multi-modal fusion neural network, feature extraction and fusion are carried out to preliminary search result, obtain multi-modal fusion feature vector, weight adjustment is carried out using attention mechanism, the correlation score of each data in preliminary search result and user search intent is calculated, according to correlation score, preliminary search result is sorted, and ordered search result is obtained;Ordered search result is input into personalized recommendation system, generates final recommendation result, based on terminal equipment and user portrait, final recommendation result is presented adaptation, determines distribution content, and is pushed to terminal equipment.But the above scheme is mainly aimed at keyword search scene, and cannot actively push according to existing data. SUMMARY

[0006] The primary technical problem to be solved by the present application is to provide a method of personalized pushing and intelligent distribution.

[0007] Another technical problem to be solved by the present application is to provide a system of personalized pushing and intelligent distribution.

[0008] To achieve the above technical purpose, the present application adopts the following technical scheme:

[0009] According to the first aspect of the embodiment of the present application, a system of personalized pushing and intelligent distribution is provided, comprising: a data acquisition module, a content module, an intelligent decision module, a distribution execution module and a monitoring optimization module.

[0010] Among them, the data acquisition module acquires user original data to form a feature vector, and sends the feature vector to the intelligent decision module;The content module sends content metadata and A / B experiment label to the intelligent decision module;The intelligent decision module processes the received data to obtain recommended content, distribution strategy and A / B experiment label, and sends them to the distribution execution module;The distribution execution module displays the received data to the user and acquires the feedback information of the user, and sends the feedback information of the user to the monitoring optimization module;The monitoring optimization module analyzes the feedback information of the user and sends the analysis result to the intelligent decision module;The intelligent decision module processes again according to the analysis result, feature vector data, content metadata and A / B experiment label, obtains new recommended content, distribution strategy and A / B experiment label, and sends them to the distribution execution module, forms a cycle of decision-making, execution, feedback and optimization, until the user completes the task of this round.

[0011] More preferably, the user original data at least includes user portrait data, user initial acquaintance evaluation data, context environment data, biological feedback data and physiological index data, for multi-dimensional analysis of user state, for processing by the intelligent decision module.

[0012] More preferably, the intelligent decision module includes a feature fusion unit, a BrainAuGPT unit, a strategy optimization unit, a feedback inlet and an output layer.

[0013] The feature fusion unit adopts a Transformer+BiLSTM+multimodal splicing architecture to provide a unified high-dimensional input for the subsequent BrainAuGPT unit; the BrainAuGPT unit adopts a Transformer+LoRA fine-tuning technology to realize lightweight decision-making, and sends the lightweight decision-making to the strategy optimization unit; the strategy optimization unit models the recommended strategy as a Markov decision process according to a multi-objective weight preset at a business level and a Q-learning reinforcement learning algorithm, obtains an optimized distribution strategy and recommended content, and sends them to the output layer; the feedback inlet sends the analysis result of the received user feedback information to the BrainAuGPT unit and the strategy optimization unit; and the output layer sends the recommended content, the distribution strategy and the A / B experiment label to the distribution execution module.

[0014] More preferably, in the Transformer+BiLSTM+multimodal splicing architecture, the Transformer encoding layer is responsible for processing context text data, captures long-distance semantic dependency through a self-attention mechanism, and outputs a context semantic vector; the BiLSTM time sequence layer is responsible for processing behavior time sequences, extracts time sequence features through a bidirectional recurrent network, and generates a time sequence vector; and the multimodal splicing is responsible for fusing the context semantic vector and the time sequence vector to obtain a feature vector.

[0015] More preferably, the distribution execution module is equipped with a hot deployment mechanism, which can load the latest experiment configuration when content metadata is updated or a new variant is generated, to avoid the efficiency loss of downtime release.

[0016] According to a second aspect of an embodiment of the present application, a personalized pushing and intelligent distribution method is provided, including the following steps:

[0017] S1: Construct a content metadata database, set an A / B experiment workflow, obtain an A / B experiment label, and collect user original data;

[0018] S2: Preprocess the user original data to obtain a feature vector;

[0019] S3: Feature fusion and intelligent decision-making are performed on the feature vector, content metadata, and A / B test label to obtain recommended content, distribution strategy, and A / B test label;

[0020] S4: The recommended content, distribution strategy, and A / B test label are pushed to the user end, and the user state is monitored to collect user feedback information;

[0021] S5: The user feedback information is analyzed to obtain an analysis result;

[0022] S6: Feature fusion and intelligent decision-making are performed on the feature vector, content metadata, A / B test label, and analysis result to obtain new recommended content, distribution strategy, and A / B test label;

[0023] S7: The new recommended content, distribution strategy, and A / B test label are pushed to the user end, and the user state is monitored to collect user feedback information, and steps S5-S7 are repeated until the user completes the current task;

[0024] S8: The completion of the current task is summarized to form a task report;

[0025] S9: The task report is pushed to the task initiator to enable the task initiator to understand the user completion or adjust the task plan according to the task report.

[0026] Preferably, the content metadata is obtained by the following steps:

[0027] The content module modularly disassembles the original content data to generate independently configurable content units, each unit including at least content type, theme label, and interaction form;

[0028] The GraphSAGE algorithm is used to construct a content unit association graph, with nodes representing content units and edges representing logical relationships between units, and a neighborhood sampling algorithm is used to generate content combinations;

[0029] The content metadata and experimental labels output by the content module to the intelligent decision-making module are used together with the feature vector and feedback information analysis result to generate a recommendation strategy.

[0030] Preferably, each content combination is assigned a unique A / B / C variant label, and traffic grouping is achieved through the GrowthBook algorithm.

[0031] Preferably, the A / B test workflow includes the following steps:

[0032] The content module standardizes the cutting of content to generate units, generates variant labels through a graph neural network, and completes traffic grouping through the GrowthBook algorithm;

[0033] The content variants and A / B test labels are transmitted to a distribution execution module, and the distribution of traffic is dynamically adjusted according to user portraits and context environments;

[0034] After user interaction, the distribution execution module collects multi-dimensional feedback data, associates variant labels, and sends them to a monitoring optimization module through a Kafka message queue;

[0035] The monitoring optimization module cleanses data, models causal relationships based on a DAG structure, calculates the causal effect value of the variant through a CausalML algorithm, and filters significant influencing factors through hypothesis testing;

[0036] Based on the conclusions of causal analysis, the LoRA fine-tuning of the BrainAuGPT unit is triggered, the strategy is optimized, the distribution of traffic is adjusted, and new variants are generated in combination with effective features, forming an A / B test optimization closed loop.

[0037] Compared with the prior art, the present application realizes dynamic adaptation of content to different user abilities, preferences and scenarios through multi-dimensional attribute labeling and flexible combination; realizes precise distribution, dynamic recombination and flexible reuse of content through fine-grained modular decomposition; realizes highly personalized content pushing and efficacy improvement through integration of multi-source data such as user portraits, evaluation results, real-time biological feedback, combination of BrainAuGPT unit and context-aware dynamic optimization algorithm; and realizes significant optimization of A / B test efficiency and gray release effect through real-time causal inference engine and distributed reasoning. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 For the first embodiment of the present application, a structure diagram of a personalized pushing and intelligent distribution system;

[0039] Figure 2 For the first embodiment of the present application, a structure diagram of an intelligent decision-making module;

[0040] Figure 3 For the second embodiment of the present application, a flowchart of a personalized pushing and intelligent distribution method;

[0041] Figure 4 For the second embodiment of the present application, a flowchart of an A / B test workflow. DETAILED DESCRIPTION

[0042] The technical content of the present application will be described in detail below in combination with the drawings and specific embodiments.

[0043] First embodiment

[0044] As Figure 1As shown, the first embodiment of the application provides a personalized push and intelligent distribution system, which at least includes a data acquisition module, a content module, an intelligent decision module, a distribution execution module and a monitoring optimization module.

[0045] The data acquisition module collects user raw data to form a feature vector and sends the feature vector to the intelligent decision module; the content module sends content metadata and A / B experiment labels to the intelligent decision module; the intelligent decision module processes the received data to obtain recommended content, distribution strategy and A / B experiment labels, and sends them to the distribution execution module; the distribution execution module displays the received data to the user and collects user feedback information, and sends the user feedback information to the monitoring optimization module; the monitoring optimization module analyzes the user feedback information and sends the analysis result to the intelligent decision module; the intelligent decision module processes again according to the analysis result, the feature vector data, the content metadata and the A / B experiment labels to obtain new recommended content, distribution strategy and A / B experiment labels, and sends them to the distribution execution module to form a decision-execution-feedback-optimization cycle until the user completes the task in this round.

[0046] In an embodiment of the application, the data acquisition module is one of the data input ends of the intelligent decision module, and its core function is to provide the intelligent decision module with raw data in the form of multi-modal vectors.

[0047] The user raw data at least includes user portrait data, user initial evaluation data, context environment data, biological feedback data and physiological index data.

[0048] The user portrait at least includes user attribute features (such as age, gender, region, etc.) and behavior labels (such as click preferences, collection records, etc.).

[0049] The user initial evaluation data refers to user attention test data, emotion evaluation data and quantitative score data.

[0050] The context environment data at least includes environmental parameters (such as geographic location, timestamp, etc.) and device status (such as terminal type, network condition) of the user, which are used to construct the space-time features of the interaction scene.

[0051] The biological feedback data at least includes voice emotion analysis (tone detection depression) and screen interaction log (click frequency, dwell time).

[0052] The physiological index data at least includes the heart rate and blood pressure of the user, which are used to evaluate the user's state in each stage of medical intervention.

[0053] The collected data needs to be standardized and multi-modal spliced to form a fusion vector that meets the input requirements of the intelligent decision module. The specific process is as follows:

[0054] First, the numerical data (such as physiological indicators) and categorical data (such as regional labels) are standardized and converted to ensure that the data dimensions are uniform; then, through multi-modal splicing technology, the features of different modalities are fused to generate a feature vector.

[0055] The data acquisition module is based on multi-modal data acquisition and has standardization processing as its core capability, which is the "data source" of the intelligent decision-making module. Its functions include raw data acquisition and preprocessing, ensuring that the intelligent decision-making module can achieve accurate recommendation based on real and multi-dimensional data, and continuously improving the strategy effect through a closed-loop mechanism, which reflects the basic support role of data in the intelligent decision-making process and the dynamic optimization value.

[0056] In an embodiment of the present application, the content module is the core hub of the intelligent decision-making system for implementing content structured management and A / B experiment design. Its core function is to disassemble raw content into reusable standardized units and generate diversified experimental variants through dynamic combination to provide the intelligent decision-making module with content metadata with experimental labels. The specific process is as follows:

[0057] First, the content module modularizes raw content such as text, video, and task flow to generate independently configurable content units (such as "task unit + video combination"), each of which includes at least content type, theme label, and interaction form. For example, in the education scenario, course content can be disassembled into units such as knowledge point explanation video, interactive exercises, and case analysis, and flexible recombination can be achieved through parameterized configuration. This process converts complex content into programmable basic components through standardized modeling, laying the foundation for subsequent experiment design.

[0058] Second, a 3-layer GraphSAGE algorithm is used to construct a content unit association graph, with nodes representing content units and edges representing logical relationships between units (such as sequence, complementarity, or exclusivity). Through neighborhood sampling algorithm, diversified content combinations are generated to ensure that the variant coverage rate exceeds 90%. Each generated content combination is assigned a unique A / B / C variant label (such as "A / B identification"), and traffic grouping is achieved through GrowthBook algorithm (default A group 50%, B group 50%), supporting gray release strategy (such as API gateway shunt control). The neighborhood sampling algorithm solves the efficiency bottleneck of traditional manual enumeration of variants, making high-dimensional content combination experiments possible.

[0059] Finally, the content metadata and experiment labels output by the content module to the intelligent decision-making module ("content metadata + experiment labels A / B") are combined with the feature vector and feedback information analysis results to participate in the generation of the recommendation strategy.

[0060] The content module plays a key role in the system closed loop interface of experiment design and strategy iteration: when the monitoring and optimization module determines the optimal variant through causal analysis, the module can dynamically adjust the content combination strategy (such as promoting the experiment label of high conversion variants) or trigger the generation of new variants (such as generating potential high-efficiency combinations based on the DAG structure inference), forming a complete link of "content unit modeling → variant generation → experiment verification → causal analysis → strategy optimization". The two-way linkage of content recommendation and model training is realized.

[0061] The content modularization module solves the problems of low efficiency of experiment design and insufficient variant coverage in traditional content recommendation through three core capabilities of standardized content unit construction, GNN-driven variant generation, and automated experiment labeling and shunting. Through causal feedback closed loop with the intelligent decision module and the monitoring module, the dynamic evolution of the content recommendation strategy is realized. The technical design not only improves the content production and experiment efficiency (such as variant generation accuracy > 90%), but also provides an interpretable optimization basis for the system, which is a key link connecting content supply and user demand in the intelligent decision-making process.

[0062] In an embodiment of the present application, the intelligent decision module is the core processing hub of the system, which undertakes the whole-process intelligent decision-making tasks of multi-modal data fusion, recommendation strategy generation, experiment optimization and dynamic feedback. The architecture design and technical implementation of the module directly determine the accuracy, real-time performance and business adaptation ability of the system. The module takes "multi-source data input - deep feature modeling - generative decision - causal closed loop optimization" as the main line, integrates five core units and a two-way data link, forms a complete intelligent decision-making system, and refers to Figure 2 , including: a feature fusion unit, a BrainAuGPT unit, a strategy optimization unit, a feedback portal and an output layer.

[0063] Among them, the feature fusion unit is the core component of multi-modal data deep modeling, which adopts a hybrid architecture of "Transformer + BiLSTM + multi-modal splicing": the Transformer encoding layer is responsible for processing context text data (such as content description, user comments), capturing long-distance semantic dependencies through self-attention mechanism, and outputting context semantic vectors; the BiLSTM time series layer is aimed at behavior time series (such as user click trajectory, video viewing progress), extracts time series features through bidirectional recurrent network, and generates time series vectors; after standardization processing, the two types of vectors, user portrait features and A / B experiment labels, are spliced through a multi-modal splicing layer to generate a fusion feature vector, providing a unified high-dimensional input for the subsequent BrainAuGPT unit. The unit solves the time series alignment problem of multi-modal data through position encoding technology, and introduces knowledge graph node features to enhance the spatial correlation modeling capability.

[0064] As the core module of generative decision-making, the BrainAuGPT unit is based on a large language model architecture and uses "Transformer + LoRA fine-tuning" technology to achieve lightweight decision-making. The base model is a pre-trained GPT-like network (BrainAuGPT), which is responsible for extracting high-level semantic representations from fused features, generating candidate recommended content (such as medical solutions, learning paths), and distributing labels (priority, gray level). LoRA (Low Rank Adaptation) technology freezes most pre-trained parameters and only updates low-rank matrix parameters, achieving memory saving while supporting online incremental training. When the feedback portal receives significant causal signals or real-time indicators break through the threshold, it automatically triggers LoRA parameter updates, and offline weight training based on CausalML is performed every 4 weeks. The output of this unit, "content recommendation + distribution label", constitutes the basic input for strategy optimization.

[0065] The strategy optimization unit achieves decision calibration through multi-objective driving, with core functions including weight setting and algorithm optimization. On the business level, multiple objective weights are preset (e.g., efficacy weight 0.4, adherence 0.35, experience 0.25 in medical scenarios), and a multi-objective optimizer (weighted sum + Pareto frontier search) is used to balance conflicting indicators. On the technical level, the Q-learning reinforcement learning algorithm is used to model the recommended strategy as a Markov decision process, with user feedback as the reward signal to optimize the distribution path and A / B traffic allocation. CausalML estimation results are also introduced to adjust the optimization direction and avoid strategy bias caused by relying solely on statistical correlation. The final output is recommendation ranking, A / B strategy routing, and priority configuration table, ensuring that the recommended content is highly aligned with business goals.

[0066] The feedback portal is the hub for interaction between the intelligent decision-making module and the monitoring optimization module, responsible for receiving, classifying, and distributing three types of feedback signals: real-time behavior feedback (completion rate, viewing time) is transmitted in real time through Kafka (delay < 10ms), causal analysis results are represented in DAG structure to represent variant influence (such as node-edge defined causal paths), and incremental training trigger signals are generated based on statistical significance. Signal processing follows the "core priority" principle: 70% of feedback is directed to the BrainAuGPT unit, triggering LoRA parameter updates; 30% is directed to the strategy optimization unit, adjusting multi-objective weights or Q-learning strategies. This hierarchical processing mechanism ensures efficient coordination between model parameter updates and strategy adjustments.

[0067] The output layer standardizes the decision-making results of the intelligent decision-making module into three categories of data and transmits them to the distribution execution module: recommended content set (including content unit, metadata), distribution strategy parameters (ranking rules, gray configuration), and A / B experiment labels, i.e., "recommended content + distribution strategy + A / B experiment labels".

[0068] The intelligent decision-making module uses LoRA technology to achieve model lightweighting, with inference delay less than 50ms, meeting real-time recommendation needs; CausalML infers causal effects based on the DAG structure to eliminate interference from confounding factors.

[0069] The value of the intelligent decision-making module also lies in its development of a "data-driven, causally explainable, and continuously evolving" intelligent decision-making paradigm: breaking down data barriers through multimodal feature fusion, avoiding the "correlation trap" through causal feedback, and improving the efficiency of strategy iteration through dynamic experimental design, ultimately achieving breakthroughs in recommendation accuracy, real-time responsiveness, and business scalability. In the future, the integration of technologies such as federated learning and self-supervised learning will further enhance the module's applicability in scenarios such as privacy protection and cross-domain migration, driving the evolution of intelligent decision-making towards a more automated and generalizable higher-level form.

[0070] In one embodiment of the present invention, the distribution execution module is the key execution hub in the system, connecting virtual strategies with real-world user interactions. Its core function is to transform the recommendations and strategies generated by the intelligent decision-making module into actual user-side presentations, and to capture multimodal feedback data in real time to support closed-loop optimization of the system. The specific process is as follows:

[0071] First, the distribution execution module receives three types of data from the intelligent decision-making module: the recommended content set, distribution strategy parameters, and A / B experiment labels. Based on these inputs, the distribution execution module uses an automated diversion mechanism to achieve precise content delivery. It utilizes the GrowthBook algorithm to group experiments and executes a grayscale release strategy through the API gateway, ensuring that different experimental variants are displayed to target users in a predetermined proportion. Technically, the React Native framework is used to develop a cross-platform user-end application, combined with the TLS 1.3 protocol to ensure data transmission security, and to achieve content distribution latency of less than 100ms, meeting the requirements of real-time recommendation scenarios.

[0072] Secondly, during the user interaction phase, the distribution execution module undertakes the critical task of collecting multimodal feedback data: it uses tracking technology to capture user behavior data (such as completion rate, viewing time, and click-through rate) in real time, and uses wearable devices or sensors to collect physiological feedback (such as heart rate and voice emotion, with features extracted using Wav2Vec) to form "multi-dimensional feedback data." This data is transmitted to the monitoring and optimization module via the Kafka message queue with a latency of less than 10ms, providing the raw signal for subsequent CausalML analysis. It should be noted that the distribution execution module automatically associates the A / B experiment tags of the content when collecting feedback, ensuring that the monitoring module can accurately track user responses to different variants.

[0073] The distribution execution module plays a pivotal role in strategy verification and feedback in the system closed loop: on the one hand, its shunting mechanism directly determines the effectiveness of A / B experiments, and through dynamic adjustment of experimental tags, the distribution execution module can accelerate the convergence of experimental conclusions; on the other hand, real-time feedback data are subjected to causal analysis by the monitoring optimization module to form optimization suggestions to feed back to the intelligent decision-making module, triggering LoRA fine-tuning of the BrainAuGPT unit and weight adjustment of the strategy optimization module. This link of "strategy execution-feedback collection-effectiveness verification-model optimization" enables the system to continuously iterate based on real user behavior, realizing dynamic evolution of the recommendation strategy.

[0074] The distribution execution module guarantees stability in high-concurrency scenarios through lightweight architecture and elastic extension design; and ensures high availability of recommendation requests through stateless design and load balancing technology. At the same time, the module supports deep collaboration with the content modularization module, and when content metadata is updated or new variants are generated, it can load the latest experimental configuration in real time through a hot deployment mechanism, avoiding the efficiency loss of traditional solutions.

[0075] The distribution execution module, through precise shunting, low-latency distribution, multi-modal feedback collection and elastic extension capability, converts the strategy output of the intelligent decision-making module into quantifiable user value, while providing a steady stream of real data for model optimization, ultimately realizing the full-chain connection of "strategy generation-execution verification-optimization upgrade", and is a key support link for the intelligent decision-making system to realize precision and real-time.

[0076] In an embodiment of the present application, the monitoring optimization module is the core of feedback analysis and strategy calibration of the system, and undertakes the key tasks of real-time processing of multi-modal user feedback data, causal effect analysis and generation of model optimization instructions, and its functions run through the "feedback-analysis-optimization" link of the system closed loop. The specific process is as follows:

[0077] First, the monitoring optimization module collects multi-dimensional feedback data from the distribution execution module in real time through the Kafka message queue, and the data range covers user behavior indicators (completion rate, click rate, viewing time), physiological feedback (heart rate, skin conductance) and voice emotion features (voiceprint vectors extracted by Wav2Vec), forming a multi-dimensional feedback data set. After standardization, these data enter the CausalML analysis module, which models the causal relationship between user behavior and content variants based on the DAG (Directed Acyclic Graph) structure, calculates the causal effect value (E[Y|do(X)]-E[Y]) of the variants through causal inference algorithms, and selects significant influencing factors through hypothesis testing. For example, when analyzing the role of a certain video combination variant in improving user retention rate, the module will exclude mixed factors such as region and time period to ensure the causal effectiveness of the conclusion.

[0078] Secondly, based on causal analysis, the monitoring optimization module generates two types of optimization outputs: one is the causal feedback result, that is, the quantitative variant effect evaluation, which is transmitted to the "feedback entrance" of the intelligent decision-making module in the form of structured data, triggering the LoRA fine-tuning of the BrainAuGPT unit and the weight adjustment of the strategy optimization module; the other is the strategy optimization suggestion, which combines business objectives and causal analysis conclusions to generate specific operation instructions (such as "increase the B variant traffic weight to 60%" and "adjust the multi-objective optimizer center rate index weight to 0.2") to guide the intelligent decision-making module to optimize the recommended strategy and content variant combination.

[0079] The monitoring optimization module uses Kubernetes cluster to realize elastic scheduling of computing resources, combines GrowthBook experiment platform and Redis cache system to ensure analysis efficiency in high-concurrency scenarios. It upgrades the "correlation analysis" of traditional recommendation systems to "causal reasoning", explicitly models the causal path between variables through DAG structure, avoids misjudging correlation as causality, and thus makes the optimization suggestion more valuable for business guidance.

[0080] In the system closed loop, user feedback data is analyzed by the monitoring optimization module to form an interpretable optimization signal, driving the intelligent decision-making module to switch from "recommendation based on statistical rules" to "decision-making based on causal mechanisms". For example, when the causal effect of a certain content unit combination is found to be significant, the module will not only suggest increasing the traffic of that variant, but also require the BrainAuGPT unit to prioritize learning the feature patterns of that variant through "incremental training trigger" signal. This "data collection-causal analysis-model optimization" closed loop mechanism enables the system to continuously accumulate effective decision-making experience and achieve spiral improvement of the recommendation strategy.

[0081] The monitoring optimization module solves the core problem of "causality unknown and optimization lag" in traditional feedback analysis through the whole process technical design of multi-modal feedback collection, causal effect quantification, and optimization instruction generation. Its value lies not only in improving recommendation accuracy, but also in injecting "explainability" genes into the system, making every strategy adjustment have business logic support through explicit causal link analysis, and becoming a key infrastructure for the intelligent decision-making system to realize "data-driven, scientific optimization".

[0082] Second embodiment

[0083] As shown in Figure 3 Based on the above system, the second embodiment of the present application provides a personalized pushing and intelligent distribution method, which at least includes the following steps:

[0084] S1: Construct a content metadata database, set up an A / B experiment workflow, obtain A / B experiment labels, and collect user original data.

[0085] S2: Preprocessing the user's original data to obtain a feature vector.

[0086] S3: Feature fusion and intelligent decision-making on the feature vector, content metadata, and A / B test label to obtain recommended content, distribution strategy, and A / B test label.

[0087] S4: Pushing the recommended content, distribution strategy, and A / B test label to the user end and monitoring the user's state to collect user feedback information.

[0088] S5: Analyzing the user feedback information to obtain an analysis result.

[0089] S6: Feature fusion and intelligent decision-making on the feature vector, content metadata, A / B test label, and analysis result to obtain new recommended content, distribution strategy, and A / B test label.

[0090] S7: Pushing the new recommended content, distribution strategy, and A / B test label to the user end and monitoring the user's state to collect user feedback information, repeating steps S5-S7 until the user completes the current task.

[0091] S8: Summarizing the completion of the current task to form a task report.

[0092] S9: Pushing the task report to the task initiator to understand the user's completion or adjusting the task plan according to the task report.

[0093] It should be noted that A / B testing is an iterative method of randomly dividing user traffic into multiple groups (such as group A and group B), each group displaying different versions of content combinations (i.e., "variants"), collecting user feedback data (such as click-through rate, completion rate, physiological indicator changes, etc.) in real time, quantitatively evaluating the differences between variants, and determining the optimal content strategy. Figure 4 In one embodiment of the present application, the A / B test workflow in step S1 includes the following steps:

[0094] First, the content module standardizes and cuts text, video, task processes and other content to generate independent content units, each of which carries metadata such as type, theme, and interaction form. Subsequently, a content unit association graph is constructed based on a three-layer GraphSAGE graph neural network, with nodes representing units and edges representing logical relationships. A neighborhood sampling algorithm is used to generate diverse combination paths to ensure that the variant coverage rate exceeds 90%. Each combination generates a unique A / B / C variant label, and the GrowthBook algorithm is used to complete the initial traffic grouping to support grayscale release configuration. This process uses the efficient modeling capabilities of GNN to reduce the exponential complexity of traditional manual enumeration of variants to the polynomial level, significantly improving efficiency while ensuring the rationality of the business logic of the variant combination.

[0095] Next, the generated content variants and A / B tags are transmitted to the distribution execution module via the API gateway, entering the precise traffic diversion phase. The content module renders user-side content based on the ReactNative cross-platform framework and utilizes TLS 1.3 protocol for encrypted data transmission, ensuring content delivery latency is kept below 100ms. During the distribution process, the system dynamically adjusts traffic distribution based on user profiles (such as age and activity level) and context (such as regional time zone and device model). For example, 70% of the A variant traffic is allocated to new users to collect cold start feedback, while existing users are diverted according to the default ratio. The API gateway supports a phased release strategy, which can expand the coverage of new variants in stages and reduce the risk of a full release. This step utilizes GrowthBook's grouping capabilities and the gateway's real-time scheduling to achieve personalized experimental traffic control, providing a scientific grouping basis for subsequent effect comparison.

[0096] Next, after the user interacts with the content, the distribution execution module activates a full-dimensional feedback capture mechanism. Front-end tracking collects explicit data such as completion rate, click-through rate, viewing time, and page dwell time in real time. At the physiological and emotional level, wearable devices collect physiological indicators such as heart rate variability and skin conductance, and the Wav2Vec model is used to extract emotional features from voice comments. All feedback data is automatically associated with corresponding A / B variant labels to form a dataset with causal tracking identifiers. The collected data is transmitted to the monitoring and optimization module via a Kafka message queue with a latency of less than 10ms to ensure real-time performance. This step establishes a three-in-one feedback system of "behavior-physiology-language", providing a multi-dimensional chain of evidence for causal analysis.

[0097] Next, the monitoring optimization module receives feedback data, and then performs standardization cleaning and causal inference processes. First, abnormal values are removed through a data filtering layer, and numerical data is normalized by Z-score. Then, the causal relationship between user behavior, variant characteristics and confounding factors is modeled based on a DAG (Directed Acyclic Graph) structure, the causal effect value of the variant is calculated using the CausalML algorithm, and significant influencing factors are selected through hypothesis testing. This process replaces traditional correlation analysis with causal inference to avoid false causal conclusions.

[0098] Finally, based on the conclusions of causal analysis, the LoRA fine-tuning of the BrainAuGPT unit is triggered, and the model is updated in a lightweight manner through low-rank adaptive technology, which saves memory while focusing on learning the feature patterns of high-optimization variants, and simultaneously guides the strategy optimization module to adjust the multi-objective weight, so that the recommended strategy is more accurately aligned with the business objectives; At the same time, the distribution execution module adjusts the traffic distribution strategy in real time through the API gateway based on the optimization suggestions, accelerates the verification and coverage expansion of effective strategies, and the content modularization module generates new experimental variants using GNN based on significant and effective variant characteristics, starting a new round of A / B testing, forming a complete closed loop of "causal analysis-model optimization-strategy iteration-variant regeneration". This process relies on the lightweight update of LoRA technology and the efficient variant generation capability of GNN to ensure that the system can continuously iterate based on real user feedback, significantly improve the accuracy of strategy optimization direction, and ultimately realize the spiral evolution and effect improvement of the intelligent decision system.

[0099] In an embodiment of the present application, the feature fusion and intelligent decision making in step S6 include the following steps:

[0100] First, the feature fusion unit performs data preprocessing on the feature vector, content metadata and A / B test label.

[0101] Second, use the Transformer encoding layer to process the context text, capture long-distance semantic dependencies through the self-attention mechanism, and generate context semantic vectors; use the BiLSTM time series layer to process the behavior sequence data, extract the front and rear correlation features through the bidirectional recurrent network, and generate the time series feature vector.

[0102] Then, through the multi-modal splicing layer, the high-dimensional fusion vector is formed along the dimension. In order to enhance the space correlation modeling, the knowledge graph technology can be introduced to build a "user-content-environment" correlation graph, and the user ID, content ID, geographic location ID, etc. are converted into graph node embeddings, which are injected into the fusion vector together with the edge attributes, solving the problem of semantic fragmentation and insufficient spatio-temporal correlation in traditional solutions, and providing structured feature input for the subsequent decision module.

[0103] Next, the fusion vector is encoded by a self-attention mechanism to extract deep association representations of user demand and content features. A fine-tuning strategy is combined with the business scenario to update the pre-trained model. The pre-trained parameters are frozen by the low-rank adaptive (LoRA) technology, and only the low-rank matrix is adjusted to save memory while supporting online incremental training.

[0104] Next, the BrainAuGPT unit outputs a set of content recommendations and distribution labels. These recommendation results constitute the basis for strategy optimization, retaining the generation capabilities of large models while achieving rapid adaptation to specific business scenarios through LoRA technology. This ensures that the generated results meet the recommendation needs of "thousand faces", while providing a configurable and flexible space for subsequent strategy optimization.

[0105] Next, based on the pre-set multi-objective weights, the recommendation process is modeled as a Markov decision process (MDP) through the Q-learning reinforcement learning algorithm. The user feedback reward signal is used to optimize the distribution path and generate a dynamic traffic distribution scheme.

[0106] When the monitoring module infers that a variant has a significant causal effect on user retention through the DAG structure, the strategy optimization module will increase the weight of the variant-related features, avoiding strategy bias caused by relying solely on statistical correlation.

[0107] Next, after the user interacts with the recommended content, the LoRA fine-tuning mechanism is triggered based on feedback information. The low-rank matrix parameters are updated based on the feature patterns of high-optimization variants. Offline training based on full causal data is performed every 4 weeks, while supporting real-time incremental training to ensure continuous learning of the latest user behavior patterns. This phase breaks the traditional one-way pattern of "data input - model output" through causal feedback, building a closed loop of "decision - feedback - optimization". This allows the module to evolve dynamically based on real user behavior, avoiding the vicious cycle of "recommendation bias - user churn".

[0108] Finally, new recommended content, distribution strategies, and A / B test labels are generated and sent to the distribution execution module.

[0109] It should be noted that the above multiple embodiments are only examples. The technical solutions of each embodiment can be combined, and all are within the protection scope of the present application.

[0110] The above provides a detailed description of the system and method of personalized push and intelligent distribution provided by the present application. For those skilled in the art, any obvious modification made without departing from the essential content of the present application will constitute an infringement of the patent right of the present application and will bear the corresponding legal responsibility.

Claims

1. A personalized push and intelligent distribution system, characterized by include: Data collection module, content module, intelligent decision-making module, distribution execution module and monitoring optimization module; The data acquisition module collects user raw data, forms feature vectors, and sends the feature vectors to the intelligent decision-making module; the content module sends content metadata and A / B experiment labels to the intelligent decision-making module; the intelligent decision-making module processes the received data to obtain recommended content, distribution strategies, and A / B experiment labels, and sends them to the distribution execution module; the distribution execution module displays the received data to users, collects user feedback information, and sends the user feedback information to the monitoring and optimization module; the monitoring and optimization module analyzes the user feedback information and sends the analysis results to the intelligent decision-making module; The intelligent decision-making module processes the analysis results, feature vector data, content metadata and A / B experiment labels again to obtain new recommended content, distribution strategies and A / B experiment labels, and sends them to the distribution execution module, forming a decision-execution-feedback-optimization cycle until the user completes the task of this round.

2. The system according to claim 1, wherein The user original data includes at least: user portrait data, user initial evaluation data, contextual environment data, biofeedback data and physiological indicator data, which are used for multi-dimensional analysis of user status for processing by the intelligent decision-making module.

3. The system according to claim 1, wherein The intelligent decision-making module includes a feature fusion unit, a BrainAuGPT unit, a strategy optimization unit, a feedback input and an output layer; Among them, the feature fusion unit adopts the architecture of Trans former+BiLSTM+multimodal splicing to provide unified high-dimensional input for the subsequent BrainAuGPT unit; the BrainAuGPT unit adopts Transformer+LoRA fine-tuning technology to achieve lightweight decision-making, and sends the lightweight decision to the policy optimization unit; the policy optimization unit models the recommendation strategy as a Markov decision process based on the multi-objective weights preset at the business level and the Q-learning reinforcement learning algorithm, obtains the optimized distribution strategy and recommended content, and sends it to the output layer; the feedback entrance sends the analysis results of the received user feedback information to the BrainAuGPT unit and the policy optimization unit; the output layer sends the recommended content, distribution strategy and A / B experiment labels to the distribution execution module.

4. The system according to claim 3, wherein In the Transformer+BiLSTM+multimodal splicing architecture, the Transformer encoding layer is responsible for processing contextual text data, capturing long-distance semantic dependencies through the self-attention mechanism, and outputting contextual semantic vectors; the BiLSTM timing layer is responsible for processing behavioral time series, extracting timing features through a bidirectional recurrent network, and generating timing vectors; and multimodal splicing is responsible for fusing the contextual semantic vector and the timing vector to obtain a feature vector.

5. The system according to claim 1, wherein The distribution execution module is equipped with a hot deployment mechanism, which can load the latest experimental configuration when the content metadata is updated or a new variant is generated, avoiding the efficiency loss of downtime release.

6. A method for personalized push and intelligent distribution, implemented based on the system according to any one of claims 1 to 5, characterized in that The steps include: S1: Build a content metadata database, set up the A / B experiment workflow, obtain A / B experiment labels, and collect user raw data; S2: Preprocess the user's original data to obtain the feature vector; S3: Performs feature fusion and intelligent decision-making on feature vectors, content metadata, and A / B experiment labels to obtain recommended content, distribution strategies, and A / B experiment labels. S4: Push recommended content, distribution strategies, and A / B experiment tags to the user end, monitor user status, and collect user feedback information; S5: Analyze user feedback information and obtain analysis results; S6: Combine feature vectors, content metadata, A / B experiment labels, and analysis results to perform feature fusion and intelligent decision-making to obtain new recommended content, distribution strategies, and A / B experiment labels. S7: Push the new recommended content, distribution strategy, and A / B experiment tags to the user, monitor the user's status, collect user feedback, and repeat steps S5 to S7 until the user completes the task of this round. S8: Summarize the completion status of tasks in this round and form a task report; S9: Push the task report to the task initiator so that the task initiator can understand the user's completion status or adjust the task plan based on the task report.

7. The method according to claim 6, wherein Obtaining the content metadata includes the following steps: The content module modularizes the original content data to generate independent and configurable content units. Each unit includes at least content type, topic label and interaction form. The GraphSAGE algorithm is used to construct a content unit association graph, with nodes representing content units and edges representing logical relationships between units. Content combinations are generated through a neighborhood sampling algorithm. The content metadata and experimental tags output by the content module to the intelligent decision-making module, together with the feature vector and feedback information analysis results, participate in the generation of recommendation strategies.

8. The method according to claim 7, wherein Each of the content combinations is assigned a unique A / B / C variant label and traffic grouping is achieved through the GrowthBook algorithm.

9. The method according to claim 6, wherein The A / B experiment workflow includes the following steps: The content module standardizes and cuts content generation units, generates variant labels through graph neural networks, and completes traffic grouping through the GrowthBook algorithm; Content variations and A / B experiment tags are transmitted to the distribution execution module, which dynamically adjusts traffic distribution based on user profiles and context. After user interaction, the distribution execution module collects multi-dimensional feedback data, associates variant tags, and sends it to the monitoring optimization module via the Kafka message queue; The monitoring and optimization module cleans data, models causal relationships based on the DAG structure, calculates the causal effect value of the variant using the CausalML algorithm, and screens significant influencing factors through hypothesis testing; Based on the conclusions of causal analysis, the LoRA fine-tuning of the BrainAuGPT unit is triggered to optimize the strategy, adjust the traffic distribution, and generate new variants based on effective features to form an A / B experiment optimization closed loop.

Images