A cloud platform-based user demand pushing e-commerce system

Abstract

The application discloses a user demand pushing e-commerce system based on a cloud platform, and belongs to the technical field of user demand pushing, and comprises a user demand perception and real-time decision module, a full-link observability platform, an intelligent routing and resource scheduling engine, a personalized model service module and a distributed message routing and guarantee module. The user demand perception and real-time decision module adopts an end-cloud collaborative architecture, is arranged on part of user terminals, is used for collecting multi-modal user interaction behavior sequences in real time, calculating short-term interest feature vectors, rearranging and deciding a local content candidate set based on a lightweight model, and simultaneously arranging part of the cloud end based on a federal learning framework to aggregate model updates of each terminal and dynamically optimize a decision strategy, thereby forming a closed loop. On the basis of realizing user demand pushing, the application can effectively solve the information cocoon problem and can realize real-time personalized recommendation.

Description

Technical Field

[0001] This invention relates to the field of user demand push technology, and more specifically, to a cloud-based user demand push e-commerce system. Background Technology

[0002] The shift from computer science terminology to cloud computing represents a major change the industry is about to face. The emergence of various cloud platforms is one of the most important aspects of this transformation. As the name suggests, these platforms allow developers to either run their programs in the cloud, use services provided in the cloud, or both.

[0003] Most systems still rely on centralized cloud-based model training and inference, resulting in delays of seconds or even longer for real-time user behavior feedback. This makes it impossible to achieve true real-time personalization and capture fleeting shifts in user interests. Furthermore, in pursuit of short-term metrics such as click-through rates, algorithms are prone to falling into information cocoons and popularity biases, leading to homogenized recommended content and damaging long-term user experience and exploration desire. This is one of the key issues that currently plague recommendation algorithms. Secondly, at the system architecture level, the scheduling of computing, storage, and network resources is often static or based on simple rules, which cannot adapt to the dynamic and ever-changing needs of heterogeneous edge-cloud environments. This results in uneven resource utilization, cloud congestion during peak periods, idle computing power at the edge and on terminals, and high overall costs. Summary of the Invention

[0004] In view of the problems existing in the prior art, the purpose of this invention is to provide a user demand push e-commerce system based on a cloud platform. In addition to realizing the push of user demand, this invention can also effectively solve the information cocoon problem and provide real-time personalized recommendations.

[0005] To solve the above problems, the present invention adopts the following technical solution:

[0006] A cloud-based user demand push e-commerce system includes: a user demand perception and real-time decision-making module, a full-link observability platform, an intelligent routing and resource scheduling engine, a personalized model service module, and a distributed message routing and guarantee module.

[0007] The user demand perception and real-time decision-making module adopts an edge-cloud collaborative architecture. The part deployed on the user terminal is used to collect multimodal user interaction behavior sequences in real time, calculate short-term interest feature vectors, and make reordering decisions on the local content candidate set based on a lightweight model. Meanwhile, the part deployed in the cloud aggregates model updates from each terminal based on a federated learning framework and dynamically optimizes the decision-making strategy to form a closed loop.

[0008] The end-to-end observability platform is used to collect, analyze in real time, and locate the root causes of end-to-end performance, logs, and business event data from user terminal behavior collection, cloud service processing to content display, and realize the intelligent operation and maintenance process through predictive alarms and closed-loop self-healing.

[0009] The intelligent routing and resource scheduling engine is used for scheduling push tasks. It collects real-time resource status and task characteristics, and selects the optimal content generation node and delivery path for each push task through a multi-objective optimization scoring function.

[0010] The personalized model service module is used to build and evolve personalized models while protecting user privacy. It is based on federated learning and transfer learning technologies, uses a large cloud model as a base, generates lightweight sub-models for different user groups or individuals, and continuously optimizes the model through online learning and causal inference technologies to correct recommendation bias.

[0011] The distributed message routing and guarantee module is used to uniformly manage and intelligently route various business messages. Based on the semantic type and priority of the message, it selects the optimal path for delivery among multiple available channels and provides at least one delivery guarantee and global order guarantee to ensure the reliable and orderly arrival of critical business messages.

[0012] Compared with the prior art, the advantages of this invention are:

[0013] (1) This invention sinks the lightweight rearrangement model and real-time feature calculation to the terminal, so that the user's intention can be instantly perceived and transformed into the display result locally, reducing the decision delay from the second level to the millisecond level, and truly achieving what you think is what you get. More importantly, the solution does not stop at simple real-time performance, but actively identifies and corrects recommendation bias through a series of composite technologies such as multimodal emotion perception, causal inference correction, and multi-objective optimization. This enables the system to not only respond to users quickly, but also guide users to discover potential interests, break the information cocoon, promote a healthy and diverse consumption ecosystem, and achieve the unity of short-term user experience and long-term commercial value. At the same time, with the help of federated learning and meta-learning, under the premise of strictly protecting user data privacy, the continuous evolution of the model and the rapid personalization of cold-start users are achieved.

[0014] (2) This invention deeply integrates end-to-end observability, intelligent resource scheduling, reliable message communication, and personalized model services to form an organism with the ability to perceive, make decisions, execute, and evolve. The intelligent routing engine performs multi-objective and predictive dynamic scheduling based on global real-time observation data, enabling computing tasks to intelligently address each other between the cloud, edge, and terminal like a flowing stream, maximizing resource utilization and reducing costs. The end-to-end observability platform transforms operation and maintenance from passive firefighting to proactive prediction and autonomous repair through dynamic baselines, causal reasoning, and an AI operation and maintenance knowledge base. The distributed message module ensures the reliable and orderly delivery of all key instructions through semantic-aware routing and global order protection. Behind all this is the free flow of data and the mutual enhancement of decisions between modules, enabling the entire system to adapt to changes in business traffic, self-optimize resource allocation, self-heal local faults, and self-evolve service capabilities like a living organism. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a user demand push e-commerce system based on a cloud platform according to the present invention.

[0016] Figure 2 This is a schematic diagram of a user demand perception and real-time decision-making module in a cloud-based user demand push e-commerce system according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] Example:

[0019] Please see Figure 1-2 A cloud-based user demand push e-commerce system includes: a user demand perception and real-time decision-making module, a full-link observability platform, an intelligent routing and resource scheduling engine, a personalized model service module, and a distributed message routing and guarantee module.

[0020] The user demand perception and real-time decision-making module adopts an edge-cloud collaborative architecture. The part deployed on the user terminal is used to collect multimodal user interaction behavior sequences in real time, calculate short-term interest feature vectors, and make reorder decisions on the local content candidate set based on a lightweight model. Meanwhile, the part deployed in the cloud aggregates model updates from each terminal based on a federated learning framework and dynamically optimizes decision-making strategies to form a closed loop.

[0021] The end-to-end observability platform is used to collect, analyze and locate the root causes of performance, logs and business event data across the entire chain from user terminal behavior collection, cloud service processing to content display, and realize intelligent operation and maintenance processes through predictive alarms and closed-loop self-healing.

[0022] The intelligent routing and resource scheduling engine is used to schedule push tasks. It collects real-time resource status and task characteristics, and selects the optimal content generation node and delivery path for each push task through a multi-objective optimization scoring function.

[0023] The Personalized Model Service Module is used to build and evolve personalized models while protecting user privacy. It is based on federated learning and transfer learning technologies, uses a large cloud model as a foundation, generates lightweight sub-models for different user groups or individuals, and continuously optimizes the model through online learning and causal inference technologies to correct recommendation bias.

[0024] The distributed message routing and guarantee module is used to uniformly manage and intelligently route various business messages. Based on the semantic type and priority of the message, it selects the optimal path for delivery among multiple available channels and provides at least one delivery guarantee and global order guarantee to ensure the reliable and orderly arrival of critical business messages.

[0025] In a specific embodiment of the present invention, user needs are perceived and decided upon in real time at the terminal, resulting in a millisecond-level response. The end-to-end observability platform acts as the nerve center of the system, monitoring the entire process from user click to cloud computing in real time. The intelligent routing engine plays the role of an intelligent scheduling center, dynamically allocating computing tasks and push paths based on the observed real-time load and network conditions. The personalized model service module is the intelligent brain of the system, responsible for producing and evolving AI models adapted to different users. The distributed message routing and guarantee module is the highly reliable blood vessel, ensuring that all kinds of notifications and instructions are delivered. These five modules do not operate in isolation, but are tightly coupled through data flow. Observational data drives routing decisions, routing status is fed back to the observation platform, and real-time user behavior is used to optimize personalized models, forming an organic whole of self-perception, decision-making, optimization, and guarantee.

[0026] Specifically, the user demand perception and real-time decision-making module includes a cloud-based collaborative training unit, a device-side real-time behavior acquisition unit, a device-side lightweight rearrangement model unit, and a device-side real-time feature engine.

[0027] The real-time behavior acquisition unit on the device side is connected to the real-time feature engine on the device side, and is used to capture and process user interaction behavior sequences locally on the user terminal in real time.

[0028] The edge-side lightweight reordering model unit receives real-time features and candidate sets from the edge-side real-time feature engine and makes reordering decisions.

[0029] The cloud-based collaborative training unit aggregates the gradients generated by the lightweight rearranged model units of each terminal after local training based on the federated learning framework, dynamically updates and distributes the model to the edge, forming a closed loop of perception and decision optimization.

[0030] In a specific embodiment of the present invention, the on-device real-time behavior acquisition unit acts as a personal observer for the user, capturing multimodal interaction sequences such as touch, voice, and vision locally without delay. The on-device real-time feature engine acts as a short-term memory, dynamically calculating the user's instantaneous interest vector using a sliding window and decay factor, capturing interest drift in seconds. The on-device lightweight reordering model is the local decision-maker, based on an improved dual-tower model that integrates long-term profiles, real-time interests, and product context, performing millisecond-level reordering of locally cached content, directly converting user intent into display results, bypassing cloud round-trip latency. The cloud collaborative training unit is the global coach, securely aggregating the local experience of thousands of terminals through federated learning, continuously optimizing the global model and distilling and distributing it, enabling all terminals to share collective wisdom without uploading raw data.

[0031] Specifically, the edge-side real-time behavior acquisition unit includes an edge-side real-time behavior acquisition unit, a cloud-based collaborative training unit, and a full-link observability platform;

[0032] The on-device real-time behavior acquisition unit is used to collect the user's touch, voice and visual gaze information locally on the user terminal through sensors, and integrates a sentiment analysis model. When it detects that the user's speech speed is faster or the pitch is higher, it automatically marks the corresponding behavior data with high-priority sentiment tags.

[0033] The cloud-based collaborative training unit works in conjunction with the edge-side unit, employing reinforcement learning algorithms to maximize the long-term value of user interaction. It optimizes the weight coefficients and strategies for ranking decisions online. High-priority sentiment tags, as an important component of the reward function in the reinforcement learning algorithm, are used to quantify the user's emotional investment and immediate satisfaction, thereby guiding the model to learn to lean towards content that can better stimulate positive emotional responses from users.

[0034] The end-to-end observability platform receives raw behavioral data and decision logs with TraceID attached from the real-time behavior acquisition unit on the edge side. By analyzing the data, it diagnoses the end-to-end performance and business anomalies from multimodal behavior perception, edge feature calculation, model reordering decision to final content display in real time.

[0035] The diagnostic results output by the end-to-end observability platform are fed back to the cloud-based collaborative training unit in real time. The diagnostic results include at least information on the attenuation of recommendation effect caused by edge model inference delay, untimely cloud model updates, or network transmission bottlenecks. The cloud-based collaborative training unit uses this feedback to dynamically adjust the frequency of federated aggregation, the strategy of personalized model distillation, and the step size of reinforcement learning exploration, thereby achieving communication perception optimization of model distribution, updates, and decision-making strategies, forming a closed-loop collaborative system that drives decision optimization based on emotion signals and ensures optimization efficiency through end-to-end observation.

[0036] The sentiment analysis model, from the perspective of the integration of psychology and artificial intelligence, transforms non-verbal emotional signals into computable high-priority labels. These labels are not only used for real-time ranking and weighting, but also serve as cross-platform collaborative signals. The cloud-based reinforcement learning reward mechanism enables the global model's learning objectives to go beyond the traditional click-through rate (CTR), encompassing the deep dimension of user emotional experience. On the observation side, the behavioral sequences carrying sentiment labels are combined with TraceID, enabling end-to-end performance analysis to distinguish the system response performance under different emotional loads.

[0037] The reinforcement learning algorithm used by the edge-side real-time behavior acquisition unit is near-end policy optimization, and its objective function is: ,in The network parameters of the representative policy are dimensionless. For the old and new strategies in state s t Choose action a t The probability of adjusting weights and selecting a sorting strategy is dimensionless. The dimensionless estimate of the dominance function at time t is used to measure the quality of the action relative to the average level. The hyperparameter is set to 0.2, which is dimensionless, to limit the policy update magnitude and ensure training stability. This algorithm enables the system to dynamically learn the optimal real-time decision-making policy based on global user feedback.

[0038] The edge-side real-time behavior acquisition unit attaches a tracking identifier (TraceID) to the raw behavior data and reports it. The end-to-end observability platform analyzes this data to diagnose end-to-end latency and anomalies from behavior perception to rearrangement display in real time, and feeds back the diagnostic results to the cloud-based collaborative training unit to optimize model distribution and update strategies.

[0039] In a specific embodiment of the present invention, the terminal-side real-time feature engine is specifically used to maintain a sliding time window locally on the terminal and calculate the user's short-term interest feature vector in real time. The dynamic update formula for the feature vector is as follows: ,in The short-term interest feature vector at time t is dimensionless. The vector from the previous time step. Let t be the behavioral event that occurs at time t. For behavior encoding functions, The decay factor has a value range of 0-1 and is used to control the rate at which historical behavior is forgotten.

[0040] The algorithm for performing rearrangement decisions in the lightweight rearrangement model unit on the end side is based on an improved dual-tower model similarity calculation, and its core ranking score formula is: ,in This is a long-term static user profile vector, synchronized from the cloud and dimensionless. Let the embedding vector of product i be dimensionless. The defined real-time short-term interest feature vector is dimensionless. The real-time context feature vector of product i is dimensionless. Calculate cosine similarity to capture long-term preference matching. This is a lightweight multilayer perceptron used to fuse real-time interests with product context and capture dynamic matching. The weight coefficients are dimensionless and adaptively learnable, optimized by cloud-based collaborative training units through reinforcement learning. Using the Sigmoid activation function, scores are mapped to the (0,1) interval. This formula integrates long-term preferences, real-time interests, and dynamic context to achieve accurate real-time personalized ranking.

[0041] The calculation formula for the end-side decision-making effectiveness evaluation sub-unit is as follows: Where k represents the number of items displayed to the user after rearrangement, in units of 1. This is an indicator function. The value is 1 when the j-th product is clicked, and 0 otherwise. This indicator is used to quickly judge the quality of decisions on the device side. When the LocalCTR is lower than the threshold, the model is immediately rolled back or cloud intervention is requested.

[0042] The MLP fusion network employs a gating mechanism, and its gating weight g is calculated as follows: ,in This is the weight matrix. For bias terms, Representing vector concatenation, this gating mechanism adaptively determines the extent to which real-time interest features should be relied upon in the final decision. With product real-time context This enhances the interpretability and scene adaptability of the model.

[0043] Specifically, it consists of a unified data collection and governance layer, a streaming data processing and storage layer, an intelligent analysis and root cause analysis layer, a visualization and insight layer, and an operation and maintenance autonomy and decision-making layer.

[0044] The unified collection and governance layer is used to collect infrastructure, application performance, network links, user behavior and business-defined events in a unified manner through adaptive probes, non-intrusive SDKs and open protocols, and automatically attach business dimension tags to break down data silos and provide high-quality, semantically rich raw data for subsequent correlation analysis.

[0045] The streaming data processing and storage layer is used for real-time streaming computation and intelligent compression storage of massive, multimodal data collected. Its integrated dynamic baseline algorithm can establish an adaptive expected range for key performance indicators. The core formula is: It can be further optimized into a seasonal adaptive baseline. This enables accurate identification of anomalies, and by establishing a spatiotemporal correlation index across data types, it supports second-level retrieval of massive historical data.

[0046] The intelligent analysis and root cause localization layer is used for in-depth analysis of the fused data. It utilizes a graph neural network (GNN) to model the system's topology and indicator correlations, automatically identifying anomalous indicator clusters. Based on a causal inference engine, combined with Bayesian networks or structural causal models, it calculates the posterior probability that different nodes are root causes. The inference formula is as follows: Accurately pinpoint the source of the fault;

[0047] The visualization and insight layer is used to transform complex technical data into intuitive and actionable business insights. It automatically discovers and renders business topology based on tracking data, forming an end-to-end full-stack observation view, and automatically maps technical indicators to business-understandable service level targets (SLO) health scores. At the same time, for specific domains such as large language model (LLM) applications, it pre-sets domain key indicators (TTFT) and dedicated analysis views to achieve out-of-the-box in-depth observation.

[0048] The autonomous operation and decision-making layer is used to realize the automation and intelligent closed loop of the operation and maintenance process. Based on the time series prediction algorithm, it performs trend prediction of core capacity to realize predictive alarms. It integrates the large language model LLM capability to build an intelligent knowledge base. When a new alarm occurs, it automatically matches historical cases and recommends handling solutions. For standardized fault scenarios, it automatically executes the plan after obtaining authorization to realize fault self-healing and form an autonomous closed loop. The tagged data output by the unified collection and governance layer is the foundation of the streaming processing layer. The time series features and abnormal events extracted by the streaming processing layer are the input of the intelligent analysis layer. The root cause localization conclusions output by the intelligent analysis layer drive the visualization layer to focus on key issues and guide the autonomous operation and maintenance layer to take precise actions. The execution results of the autonomous layer and new business events are fed back to the collection layer to form a continuous learning and self-optimizing intelligent operation and maintenance life cycle.

[0049] In a specific embodiment of the present invention, the core reasoning formula is as follows: Where P(R|E) represents the dimensionless posterior probability of a potential root cause R occurring given the observed anomaly set E; P(E|R) represents the dimensionless likelihood probability of observing the current anomalous pattern E given the occurrence of root cause R, obtained through training from historical fault data or domain knowledge graphs; P(R) represents the dimensionless prior probability of root cause R occurring; and P(E) represents the dimensionless marginal probability of evidence E occurring. The real-time stream computing engine integrates a dynamic baseline algorithm to establish an adaptive expected range for API response time and error rate performance metrics. Its core formula is... , This represents the dynamic baseline range of the indicator at time point t, with the same units as the original indicator. This represents the moving average of the indicator values ​​within a time window w, with the same units as the original indicator. This represents the moving standard deviation of the indicator value within the same time window, with the same units as the original indicator. The configurable, dimensionless sensitivity coefficient is used to control the leniency of the alarm threshold.

[0050] In a specific embodiment of the present invention, the causal reasoning engine unit is further configured as follows:

[0051] In specific scenarios such as major e-commerce promotions, a business impact quantification model is introduced. This model correlates technical anomalies with order failures and GMV losses, and its impact assessment formula is as follows: Where Impact represents the dimensionless score of the overall business impact of this anomaly event. This indicates that the abnormal deviation of the i-th affected business indicator is in the same unit as the corresponding KPI. The preset weighting coefficient for this business indicator is dimensionless, reflecting its business importance;

[0052] The dynamic baseline algorithm is further optimized into a seasonal adaptive baseline algorithm, which can automatically detect and learn the periodic patterns of indicators. Its baseline calculation is extended as follows: ,in This represents the seasonal component at time point t, with the same units as the original indicator, calculated from historical data for the same period. It represents the trend component, with the same unit as the original indicator, and is derived from moving average or linear regression.

[0053] Specifically, the intelligent routing and resource scheduling engine includes a push perception layer, a push optimization decision layer, a heterogeneous resource execution layer, and a push guarantee layer;

[0054] The push perception layer is used to collect system status and request characteristics strongly related to push tasks in real time. The push optimization decision layer is used to select the optimal calculation and delivery path for each push task based on the optimization goals of the push system. The heterogeneous resource execution layer is used to convert scheduling decisions into actual operations on heterogeneous push resources. The push guarantee layer is used to monitor the entire link status of push tasks and intervene and recover when anomalies occur, thereby jointly ensuring low latency and high reliability of push content from cloud generation to user display.

[0055] The push perception layer specifically includes a resource status acquisition unit, a request feature parsing unit, and a topology and link discovery unit;

[0056] The resource status acquisition unit is used to collect performance indicators of the backend push content generation node in real time through a proxy or open telemetry protocol. The performance indicators include the current queue depth of the content generation service, GPU utilization, and network link latency and packet loss rate from the node to user groups in different regions.

[0057] The request feature parsing unit is used to parse the input push task in real time and extract push features specifically for scheduling decisions. Push features include user demand intensity coefficient, a list of candidate product or content IDs to be pushed and their quantity, and the version number of the lightweight reordering model currently deployed on the user terminal of the request source.

[0058] The topology and link discovery unit dynamically maintains the collected link status information to generate network topology and dependencies between service nodes, edge cache nodes, message push channels, and end users, providing a basis for computation path selection.

[0059] The push optimization decision layer includes a strategy rule base unit, a real-time scoring and scheduling unit, and a causal analysis and route optimization unit.

[0060] The strategy rule base unit is used to store predefined scheduling rules for push scenarios, prioritize the allocation of high-performance generation nodes to users with high demand intensity, or route requests for specific model versions to service nodes that are compatible with that model version.

[0061] The real-time scoring and scheduling unit calculates a comprehensive score for each candidate push content generation node or delivery path. Its core push optimization scoring function trade-off formula is as follows: ,in The function estimates the total latency from task assignment, content generation, network transmission to on-device display. To evaluate whether the scheduling decision can effectively utilize the cached content of user terminals or edge nodes and avoid repeated generation and transmission, The function calculates the economic cost of using the resource node. The weights are dynamically configurable and are used to strike a balance between minimizing push latency, maximizing cache utilization, and controlling costs.

[0062] The Causal Analysis and Routing Optimization Unit is used to analyze the delay, loss, or failure of push content delivery caused by improper routing strategies. When such push anomalies are detected, it locates the nodes or links that may cause the problem based on a simple rule engine or statistical model, and dynamically adjusts the weights in the scoring function or directly modifies the routing table to avoid faulty nodes and switch subsequent traffic to a better path.

[0063] The real-time scoring and scheduling unit calculates... In addition to considering the instantaneous latency of the current network, an estimate of the processing time of the content generation service is also introduced. This estimate is based on the number of candidate content in the request characteristics and the user demand intensity coefficient, and is calculated linearly or by looking up a table to predict the total latency and scheduling.

[0064] The heterogeneous resource execution layer specifically includes a unified resource adaptation unit, a scheduling instruction execution unit, and a push traffic distribution unit;

[0065] The unified resource adaptation unit is used to abstract different AIGC services, traditional recommendation ranking services, edge caching services, vendor channels, and SMS gateways into a unified push resource node, and encapsulate their calling interfaces;

[0066] The scheduling instruction execution unit is responsible for converting the target resource node information output by the push optimization decision layer into specific API call commands, calling the generation interface of a specific AIGC service, or sending tasks to a specified message push service.

[0067] The push traffic distribution unit is used to perform load balancing among multiple AIGC service instances based on the decision results, and distribute push tasks to them for execution;

[0068] The push guarantee layer specifically includes a push link monitoring unit, an automatic fault tolerance and retry unit, and a strategy effectiveness evaluation unit;

[0069] The push link monitoring unit tracks the status of push tasks in real time, monitoring end-to-end latency and success rate from scheduling decisions, content generation, network transmission to end-side confirmation;

[0070] The automatic fault tolerance and retry unit is used to automatically replace another generation node to re-execute the task or switch to a backup push channel when it is detected that the push task is unresponsive in the content generation service or the message push channel returns a failure.

[0071] The strategy performance evaluation unit is used to continuously compare the average latency and cache hit rate under different scheduling strategies. When a new strategy causes the metrics to deteriorate, it automatically rolls back to the previous stable strategy.

[0072] In a specific embodiment of the present invention, the core multi-objective optimization scoring function of the real-time scoring and scheduling algorithm unit is: ,in The overall score for resource instance i is dimensionless. This is a performance scoring function. The input is the real-time performance metric of instance i, and the output is a dimensionless normalized score, the unit of which depends on the specific metric. This is a cost scoring function. The input is the estimated cost of use case i, and the output is a dimensionless normalized score. The cost unit is either currency or virtual points. For the Service Level Agreement (SLA) matching function, evaluate the degree of fit between the capabilities of instance i and the SLA requirements of request j, and output a normalized score that is dimensionless. The weighting coefficients for performance, cost, and SLA are dimensionless and can be dynamically adjusted to adapt to different optimization objectives.

[0073] The business impact quantification decision-making formula for the heterogeneous resource execution layer can be expressed as follows: Where Adjustment is the adjustment action of the final decision, P(R|E) is the dimensionless probability of the root cause R occurring under the condition that abnormal evidence E is observed, and Impact(R) is the dimensionless estimated impact of the root cause R on core business indicators.

[0074] Push guarantee layer is represented as , where t is the current time, Δt is the forecast window, and History is the historical sequence data.

[0075] Specifically, the push perception layer and the push optimization decision layer work together to achieve fine-grained scheduling of LLM perception for large language model LLM inference scenarios:

[0076] The number of Prompt Tokens and the estimated number of Output Tokens retrieved from the push perception layer are calculated.

[0077] Based on this characteristic, the push optimization decision layer prioritizes scheduling computationally intensive requests with high Prompt Tokens to high-performance GPU clusters, and schedules latency-sensitive requests with high Output Tokens to high-memory-bandwidth or low-load GPU clusters, thereby achieving optimized scheduling by separating the computation stages.

[0078] In a specific embodiment of the present invention, computation-aware scheduling of LLM inference requests is implemented. The request feature parsing unit meticulously analyzes the composition of each inference request, extracting two key features: Prompt Tokens and estimated Output Tokens. The intelligent decision layer then performs differentiated routing based on these features, prioritizing requests with high Prompt Tokens to high-performance GPU clusters and scheduling requests with high Output Tokens to dedicated clusters. This essentially achieves computational separation and resource adaptation between the Prefill and Decode stages in LLM inference.

[0079] Specifically, the personalized model service module includes a unified multi-source feature engineering and privacy computing layer, a personalized model factory and lifecycle management layer, a highly available model service and inference layer, and an online learning and policy evolution layer;

[0080] The unified multi-source feature engineering and privacy computing layer specifically includes a multimodal user behavior perception unit, a contextual environment fusion unit, and a privacy-preserving feature extraction unit;

[0081] The multimodal user behavior perception unit is used to collect multi-source interaction data of users on the terminal in real time, including text clicks, page dwell time, voice search commands and visual attention areas obtained through compliant means, and encode unstructured data into a unified feature vector;

[0082] The context fusion unit is used to acquire and fuse the real-time context when the request occurs, including geographical location, time, device type, network status and current promotional activity information, to generate an context vector;

[0083] The privacy-preserving feature extraction unit employs federated feature engineering and differential privacy technology to complete the feature extraction and desensitization of the original behavioral data locally on the user terminal or edge node, and only uploads the encrypted or noisy feature summary to the cloud to ensure that the user's original data does not leave the domain.

[0084] In a specific embodiment of this invention, a personalized model production line that protects privacy, continuously evolves, and corrects causal biases is constructed. A unified multi-source feature engineering and privacy computing layer processes raw data at the terminal or edge using federated feature engineering and differential privacy technology, ensuring data remains within the domain and protecting privacy from the source. The personalized model factory and lifecycle management layer are the core production line. User dynamic clustering units divide interest groups in real time, while the model factory unit, based on a large cloud model, rapidly fine-tunes group or individual models using transfer learning and meta-learning techniques, and adapts to edge deployment through knowledge distillation and compression. The federated aggregation and update unit periodically and securely aggregates terminal model updates, driving global model evolution. The high-availability model service and inference layer are responsible for agile model deployment and gray-scale verification. The online learning and strategy evolution layer balances business, experience, and system goals through multi-objective optimization algorithms, and uses causal inference and correction units to identify and correct popularity and position biases in the recommendation system, striving to restore users' true preferences.

[0085] Specifically, the personalized model factory and lifecycle management layer includes user dynamic clustering units, model factory units, and federated aggregation and update units;

[0086] The user dynamic clustering unit is based on real-time user feature vectors and long-term interest profiles. It uses an online clustering algorithm to dynamically divide users into different interest groups, providing a foundation for the construction of a group-based personalized model.

[0087] The model factory unit is responsible for the full lifecycle management of personalized models. Based on user clustering results or individual user data, it uses transfer learning technology and cloud-pre-trained large models as a base to quickly fine-tune and generate lightweight sub-models that are adapted to different groups or individuals. At the same time, it uses incremental learning and knowledge distillation algorithms to continuously optimize the model and compress its size to adapt to edge deployment.

[0088] The federated aggregation and update unit is based on the federated learning framework. It periodically aggregates local model updates from a large number of terminal devices, generates a global model improvement version, and then securely distributes it to each terminal, so as to enable all user data to jointly optimize the model without leaving the local device.

[0089] The highly available model service and inference layer includes a model repository and version management unit, a dynamic loading and inference engine unit, and an A / B testing and performance evaluation unit;

[0090] The model repository and version management unit stores and manages personalized models of different versions and types, supporting rapid model rollback and canary releases. The dynamic loading and inference engine unit loads the corresponding personalized model from the model repository into memory in real time based on the features and user ID carried in the user request, providing high-concurrency, low-latency online inference services. It supports the deployment of suitable lightweight models to user terminals for local inference. The A / B testing and performance evaluation unit allocates traffic for different model strategies or parameter configurations and collects key business indicators through real-time tracking to scientifically evaluate model performance and support rapid iteration.

[0091] The online learning and strategy evolution layer includes a real-time feedback learning unit, a multi-objective optimization and game theory unit, and a causal inference and correction unit.

[0092] The real-time feedback learning unit is used to take the user's real-time feedback on the recommendation results as a reinforcement learning signal, and adjust the feature weights or recall strategies in the ranking model in real time through online learning algorithms.

[0093] The multi-objective optimization and game theory unit is used to employ multi-objective optimization algorithms to balance business objectives, user experience objectives, and system objectives.

[0094] The causal inference and correction unit uses a causal inference model to identify and correct biases in the recommendation system, and employs the inverse probability weighted loss function (IPW) to correct the bias. ,in The corrected loss value is dimensionless, where N is the total number of samples. The indicator function is 1 when the i-th item is observed by the user, and 0 otherwise. The estimated probability of the i-th item being observed is dimensionless and is learned from the exposure model. The original loss function, such as cross-entropy loss, is dimensionless;

[0095] When generating personalized models, the model factory unit introduces a meta-learning mechanism. This mechanism pre-trains a meta-model in the cloud using a large amount of user data. This meta-model can quickly adapt and generate personalized model parameters for the target user based on a small amount of new behavioral data. The core optimization objective is... ,in The parameters of the metamodel are dimensionless. To distribute tasks One of the tasks in sampling, For the task The loss function on is dimensionless. For the task The parameters of the personalized model after rapid adaptation are dimensionless. Adapts to dimensionless learning rates;

[0096] The multimodal user behavior perception unit works in collaboration with the causal inference and correction unit to achieve a deeper level of modeling of users' true intentions and satisfaction. The multimodal data provided by the perception unit serves as an instrumental variable or confounding factor in causal inference, used to more accurately estimate users' intrinsic preferences for products and remove biases caused by the external display environment, thereby generating a purer and more accurate personalized model.

[0097] In a specific embodiment of the present invention, the core aggregation formula FedAvg of the federated aggregation and update unit is: ,in Let K represent the dimensionless global model parameters after aggregation in round t+1, and let K represent the total number of clients participating in this training round. This represents the number of samples in the local dataset of the k-th client, and n represents the total number of samples from all participating clients. , This represents the model parameters of the k-th client after local training in round t+1;

[0098] The multi-objective loss function of the multi-objective optimization and game unit is as follows: ,in The total loss value is dimensionless. Click-through rate prediction error loss, such as cross-entropy loss, is dimensionless. Conversion rate prediction error loss is dimensionless. The diversity penalty for the recommended list can be represented by a dimensionless negative value for the similarity of products within the list. The novelty penalty term is used to encourage the recommendation of new products that users have not seen before. It can be represented by a dimensionless function of the negative correlation between the product's historical exposure frequency and its novelty. The weighting coefficients for each objective are dimensionless and can be dynamically adjusted through grid search or Bayesian optimization.

[0099] The correction formula is ,in The corrected loss value is dimensionless, where N is the total number of samples. The indicator function is 1 when the i-th item is observed by the user, and 0 otherwise. The estimated probability of the i-th item being observed is dimensionless and is learned from the exposure model. The original loss function, such as cross-entropy loss, is dimensionless;

[0100] The core optimization objective of the model factory unit is ,in The parameters of the metamodel are dimensionless. To distribute tasks One of the tasks in sampling, For the task The loss function on is dimensionless. For the task The parameters of the personalized model after rapid adaptation are dimensionless. The learning rate is dimensionless.

[0101] Specifically, the distributed message routing and guarantee module includes a unified message access and semantic parsing layer, an intelligent routing decision and scheduling layer, a reliable delivery and order guarantee layer, and an autonomous operation and maintenance and elastic scaling layer.

[0102] The unified message access and semantic parsing layer includes a multi-protocol adaptation unit, a message semantic parsing and classification unit, and a unified message encapsulation unit;

[0103] The multi-protocol adaptation unit supports multiple message access protocols, including HTTP / 2, gRPC, MQTT, AMQP, and custom binary protocols, and enables transparent conversion between protocols. At the same time, the unit has a built-in connection pool management and session persistence mechanism to cope with the concurrent connections of a large number of terminal devices.

[0104] The message semantic parsing and classification unit is used to perform real-time parsing of incoming message payloads, extract message type, business ID, priority, and timeliness requirement fields, and classify messages into real-time command type, reliable transaction type, order sensitive type, or batch throughput type according to preset rules or machine learning models, providing a basis for subsequent differentiated routing.

[0105] The unified message encapsulation unit is used to encapsulate messages from different sources and formats into internal message objects with a unified header structure;

[0106] The intelligent routing decision and scheduling layer includes a global topology and state awareness unit, a multi-factor routing scoring and decision unit, and a predictive routing and load balancing unit.

[0107] The global topology and status awareness unit collects and maintains the health status, current load, network link quality, and geographical location information of all routing nodes in the system in real time, and constructs a dynamic global resource map. The multi-factor routing scoring and decision unit calculates a comprehensive score for each candidate routing path or node based on the perceived global status. The predictive routing and load balancing unit integrates the time series prediction model LSTM to predict the future load trend of key nodes and links, and performs forward-looking traffic scheduling based on the prediction results.

[0108] The reliable delivery and order guarantee layer includes a delivery state machine and persistence unit, a at least-once delivery guarantee unit, and a global order guarantee unit;

[0109] The delivery state machine and persistence unit maintain a state machine for each message and persists the critical states to a reliable distributed storage to ensure that the message state can be recovered in the event of a node failure.

[0110] The at-last-delivery guarantee unit is implemented based on the acknowledgment and retransmission mechanism. When a message is marked as a reliable transaction, the system ensures that it is consumed at least once. The core mechanism of the at-last-delivery guarantee unit includes the producer waiting for the Broker to confirm after sending the message, the Broker confirming to the producer after persisting the message, and the consumer confirming to the Broker after successfully processing the message. If no confirmation is received, the process is retried according to the exponential backoff algorithm.

[0111] The global order guarantee unit provides global order guarantee across nodes and partitions for order-sensitive messages. Through a global order coordinator, it assigns monotonically increasing sequence numbers to messages under the same business ID and ensures that consumers process them strictly in the order of the sequence numbers. Its order conflict detection and resolution algorithm uses a vector clock for version comparison. When out-of-order delivery occurs, it triggers a mechanism to delay delivery or wait for the arrival of the preceding message.

[0112] The autonomous operation and maintenance and elastic scaling layer includes an intelligent monitoring and root cause analysis (RCA) unit, a self-healing and elastic scaling engine unit, and a cost and performance optimization unit.

[0113] The intelligent monitoring and root cause analysis (RCA) unit monitors SLA metrics such as message backlog, end-to-end latency, and error rate in real time, and automatically analyzes the root causes of failures based on the rule engine and topology dependencies.

[0114] The self-healing and elastic scaling engine unit automatically executes predefined repair actions based on monitoring and RCA results. When a message queue is detected to be continuously piling up, it automatically expands the consumer instances of that queue. When a routing node is marked as unhealthy, it automatically isolates it from the routing table and switches traffic to a backup node.

[0115] The cost and performance optimization unit continuously analyzes the cost-effectiveness of different routing paths. Under the premise of meeting the SLA, it automatically schedules low-priority messages with high timeliness tolerance to lower-cost paths or resources, thereby optimizing the total cost of ownership (TCO).

[0116] In a specific embodiment of the present invention, the header of the internal message object includes a globally unique message ID, a business category label, a priority weight, a creation timestamp, an expected delivery deadline, and traceability information.

[0117] The formula for triggering elastic scaling decisions is as follows: , where α and β are configurable, dimensionless threshold coefficients;

[0118] The multi-factor routing scoring and decision-making unit further integrates a dynamic weight tuning subunit based on reinforcement learning (RL). This subunit models routing decisions as a Markov decision process (MDP), using the overall message delivery success rate, average latency, and unit cost as reward signals, and learns online to dynamically adjust the weight coefficients in the scoring function. This enables the system to adapt to constantly changing business models and resource environments;

[0119] The message semantic parsing and classification unit works in collaboration with the global order guarantee unit to achieve deep perception and guarantee of the order of business logic. The parsing unit not only identifies message types, but also extracts business objects and state transition dependencies. The order guarantee unit uses this dependency graph, rather than a simple FIFO queue, to arbitrate and guarantee the processing order of messages, thereby solving the problem of sequential consistency in complex business processes across multiple asynchronous services.

[0120] The global topology and status awareness unit can identify and manage nodes deployed in different cloud service providers and private data centers. The intelligent routing decision unit can realize intelligent scheduling and disaster avoidance of messages in multiple clouds according to policies. For example, when a failure occurs in a certain cloud area, all traffic can be seamlessly switched to other healthy areas.

[0121] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A user demand push e-commerce system based on a cloud platform, characterized in that, include: The module includes a user demand perception and real-time decision-making module, a full-link observability platform, an intelligent routing and resource scheduling engine, a personalized model service module, and a distributed message routing and guarantee module. The user demand perception and real-time decision-making module adopts an edge-cloud collaborative architecture. The part deployed on the user terminal is used to collect multimodal user interaction behavior sequences in real time, calculate short-term interest feature vectors, and make reordering decisions on the local content candidate set based on a lightweight model. Meanwhile, the part deployed in the cloud aggregates model updates from each terminal based on a federated learning framework and dynamically optimizes the decision-making strategy to form a closed loop. The end-to-end observability platform is used to collect, analyze in real time, and locate the root causes of end-to-end performance, logs, and business event data from user terminal behavior collection, cloud service processing to content display, and realize the intelligent operation and maintenance process through predictive alarms and closed-loop self-healing. The intelligent routing and resource scheduling engine is used for scheduling push tasks. It collects real-time resource status and task characteristics, and selects the optimal content generation node and delivery path for each push task through a multi-objective optimization scoring function. The personalized model service module is used to build and evolve personalized models while protecting user privacy. It is based on federated learning and transfer learning technologies, uses a large cloud model as a base, generates lightweight sub-models for different user groups or individuals, and continuously optimizes the model through online learning and causal inference technologies to correct recommendation bias. The distributed message routing and guarantee module is used to uniformly manage and intelligently route various business messages. Based on the semantic type and priority of the message, it selects the optimal path for delivery among multiple available channels and provides at least one delivery guarantee and global order guarantee to ensure the reliable and orderly arrival of critical business messages.

2. The user demand push e-commerce system based on a cloud platform according to claim 1, characterized in that, The user demand perception and real-time decision-making module includes a cloud-based collaborative training unit, a device-side real-time behavior acquisition unit, a device-side lightweight rearrangement model unit, and a device-side real-time feature engine. The cloud-based collaborative training unit aggregates the gradients generated by the lightweight rearranged model units of each terminal after local training based on the federated learning framework, dynamically updates and distributes the model to the edge, forming a closed loop of perception and decision optimization. The terminal-side real-time behavior acquisition unit is connected to the terminal-side real-time feature engine and is used to capture and process user interaction behavior sequences locally on the user terminal in real time. The edge-side lightweight reordering model unit receives real-time features and candidate sets from the edge-side real-time feature engine and makes reordering decisions. The on-device real-time feature engine is specifically used to maintain a sliding time window locally on the terminal and calculate the user's short-term interest feature vector in real time.

3. The user demand push e-commerce system based on a cloud platform according to claim 2, characterized in that, The edge-side real-time behavior acquisition unit specifically includes an edge-side real-time behavior acquisition unit, a cloud-based collaborative training unit, and a full-link observability platform; The terminal-side real-time behavior acquisition unit is used to collect the user's touch, voice and visual gaze information locally on the user terminal through sensors, and integrates an emotion analysis model. When it detects that the user's speech speed is faster or the pitch is higher, it automatically marks the corresponding behavior data with a high-priority emotion tag. The cloud-based collaborative training unit works in collaboration with the edge unit, employing a reinforcement learning algorithm to maximize the long-term interaction value of users. It optimizes the weight coefficients and strategies of ranking decisions online. The high-priority sentiment tags, as an important component of the reward function in the reinforcement learning algorithm, are used to quantify the user's emotional investment and immediate satisfaction, thereby guiding the model to learn to lean towards content that can better stimulate the user's positive emotional response. The end-to-end observability platform receives raw behavioral data and decision logs with TraceID appended from the end-side real-time behavior acquisition unit. By analyzing the data, it diagnoses the end-to-end performance and business anomalies from multimodal behavior perception, end-side feature calculation, model reordering decision to final content display in real time. The diagnostic results output by the end-to-end observability platform are fed back to the cloud-based collaborative training unit in real time. The diagnostic results include at least information on the attenuation of recommendation effect caused by edge model inference delay, untimely cloud model update or network transmission bottleneck. The cloud-based collaborative training unit uses this feedback to dynamically adjust the frequency of federated aggregation, the strategy of personalized model distillation and the step size of reinforcement learning exploration, thereby realizing the communication perception optimization of model distribution, update and decision-making strategies, forming a closed-loop collaborative system that drives decision optimization with emotion signals and ensures optimization efficiency through end-to-end observation. The sentiment analysis model, from the perspective of integrating psychology and artificial intelligence, transforms non-verbal emotional signals into computable high-priority labels. These labels are not only used for real-time ranking and weighting but also serve as cross-platform collaborative signals. The model incorporates a reinforcement learning reward mechanism in the cloud, enabling the global model's learning objective to surpass the traditional click-through rate (CTR) and encompass the deeper dimension of user emotional experience. On the observation side, behavioral sequences carrying sentiment labels are combined with TraceID, allowing end-to-end performance analysis to distinguish system response performance under different emotional loads.

4. The user demand push e-commerce system based on a cloud platform according to claim 1, characterized in that, The intelligent routing and resource scheduling engine includes a push perception layer, a push optimization decision layer, a heterogeneous resource execution layer, and a push guarantee layer. The push perception layer is used to collect system status and request characteristics strongly related to push tasks in real time. While collecting general server performance indicators, it also collects user demand intensity coefficients, a list of candidate product or content IDs to be pushed and their quantity, and the version number of the lightweight reordering model currently deployed on the user terminal of the request source, providing multi-dimensional decision-making basis for subsequent refined and scenario-based scheduling. The push optimization decision layer is used to select the optimal computation and delivery path for each push task based on the specific optimization objectives of the push system. It adopts a push optimization scoring function to simultaneously weigh three often conflicting objectives: minimizing the total latency from task delivery to end-side display, maximizing the cache hit rate of user terminals or edge nodes, and controlling resource usage costs. It achieves multi-objective joint optimization through dynamically configurable weights, rather than traditional single-objective load balancing. The heterogeneous resource execution layer is used to convert scheduling decisions into actual operations on heterogeneous push resources. Through a unified resource abstraction model, it encapsulates heterogeneous resources such as different AIGC services, traditional recommendation ranking services, edge caching services, vendor push channels and SMS gateways into push resource nodes with standardized interfaces, shielding underlying differences and achieving unified and low-latency execution of scheduling instructions. The push guarantee layer is used to monitor the entire link status of push tasks and intervene and recover when anomalies occur. It constructs a closed-loop guarantee system that includes real-time monitoring, automatic fault tolerance and policy evaluation. When push failure or delay anomaly is detected, it can automatically trigger retry actions such as changing nodes or switching channels, and realize automatic rollback of scheduling policies based on policy effect comparison to ensure the ultimate reliability of push services. The features and states collected by the push perception layer are the basis for the push optimization decision layer to score and make decisions. The output instructions of the push optimization decision layer are executed by the heterogeneous resource execution layer. The push guarantee layer monitors the task status of the execution layer and feeds back any anomalies it finds to the causal analysis and routing optimization unit of the push optimization decision layer, driving it to dynamically adjust the scoring weights or routing strategies.

5. The user demand push e-commerce system based on a cloud platform according to claim 4, characterized in that, The push perception layer and the push optimization decision layer work together to achieve fine-grained scheduling of LLM perception for large language model LLM inference scenarios: The number of Prompt Tokens and the estimated number of Output Tokens retrieved from the push perception layer are calculated. Based on this characteristic, the push optimization decision layer prioritizes scheduling computationally intensive requests with high Prompt Tokens to high-performance GPU clusters, and schedules latency-sensitive requests with high Output Tokens to high-memory-bandwidth or low-load GPU clusters, thereby achieving optimized scheduling by separating the computation stages.

6. The user demand push e-commerce system based on a cloud platform according to claim 4, characterized in that, The push perception layer specifically includes a resource status acquisition unit, a request feature parsing unit, and a topology and link discovery unit; The resource status acquisition unit is used to collect performance indicators of the backend push content generation node in real time through a proxy or open telemetry protocol. The performance indicators include the current queue depth of the content generation service, GPU utilization, and network link latency and packet loss rate from the node to user groups in different regions. The request feature parsing unit is used to parse the input push task in real time and extract push features specifically for scheduling decisions. The push features include the user demand intensity coefficient, the list of candidate product or content IDs to be pushed and their quantity, and the version number of the lightweight rearrangement model currently deployed on the user terminal of the request source. The topology and link discovery unit generates network topology and dependencies between service nodes, edge cache nodes, message push channels, and end users based on the collected link status dynamic maintenance content, providing a basis for computation path selection.

7. A user demand push e-commerce system based on a cloud platform according to claim 4, characterized in that, The push optimization decision layer includes a strategy rule base unit, a real-time scoring and scheduling unit, and a causal analysis and routing optimization unit. The strategy rule base unit is used to store predefined scheduling rules for push scenarios, prioritize the allocation of high-performance generation nodes to users with high demand intensity, or route requests for specific model versions to service nodes that are compatible with that model version. The real-time scoring and scheduling unit is used to calculate a comprehensive score for each candidate push content generation node or delivery path. Its core push optimization scoring function trade-off formula is as follows: ,in The function estimates the total latency from task assignment, content generation, network transmission to on-device display. To evaluate whether the scheduling decision can effectively utilize the cached content of user terminals or edge nodes and avoid repeated generation and transmission, The function calculates the economic cost of using the resource node. The weights are dynamically configurable and are used to strike a balance between minimizing push latency, maximizing cache utilization, and controlling costs. The causal analysis and routing optimization unit is used to analyze the delay, loss or failure of push content delivery caused by improper routing strategies. When such push anomalies are detected, the unit locates the nodes or links that may cause the problem based on a simple rule engine or statistical model, and dynamically adjusts the weights in the scoring function or directly modifies the routing table to avoid faulty nodes and switch subsequent traffic to a better path.

8. A user demand push e-commerce system based on a cloud platform according to claim 4, characterized in that, The real-time scoring and scheduling unit calculates... In addition to considering the instantaneous latency of the current network, the system also introduces an estimate of the processing time of the content generation service. This estimate is based on the number of candidate content in the request characteristics and the user demand intensity coefficient, and is calculated linearly or by looking up a table to predict the total latency and scheduling.

9. A user demand push e-commerce system based on a cloud platform according to claim 4, characterized in that, The heterogeneous resource execution layer specifically includes a unified resource adaptation unit, a scheduling instruction execution unit, and a push traffic distribution unit; The unified resource adaptation unit is used to abstract different AIGC services, traditional recommendation ranking services, edge caching services, vendor channels, and SMS gateways into a unified push resource node, and encapsulate their calling interfaces; The scheduling instruction execution unit is responsible for converting the target resource node information output by the push optimization decision layer into specific API call commands, calling the generation interface of a specific AIGC service, or sending tasks to a specified message push service. The push traffic distribution unit is used to perform load balancing among multiple AIGC service instances based on the decision results, and distribute push tasks to them for execution.

10. A user demand push e-commerce system based on a cloud platform according to claim 4, characterized in that, The push protection layer specifically includes a push link monitoring unit, an automatic fault tolerance and retry unit, and a strategy effect evaluation unit; The push link monitoring unit tracks the status of push tasks in real time, and monitors the end-to-end latency and success rate from scheduling decision, content generation, network transmission to end-side confirmation. The automatic fault tolerance and retry unit is used to automatically replace another generation node to re-execute the task or switch to a backup push channel for delivery when it is detected that the push task is unresponsive in the content generation service or the message push channel returns a failure. The strategy effect evaluation unit is used to continuously compare the average latency and cache hit rate under different scheduling strategies. When a new strategy causes the indicators to deteriorate, it automatically rolls back to the previous stable strategy.

Images