Intelligent recommendation method based on large model fusion computing network service

Abstract

The invention discloses an intelligent recommendation method based on a large model fusion computing network service, and relates to the field of artificial intelligence and large model application crossing. Comprising the steps of 1, converting an original demand input by a user into a structured computing network demand index, 2, collecting and integrating state information of cloud resources and network resources in real time, and providing dynamic computing network resource data for large model reasoning, 3, carrying out deep matching on the user demand and the computing network resources, generating a candidate scheme set of multi-objective optimization, and carrying out multi-objective optimization according to the candidate scheme set of the multi-objective optimization. 4, optimizing a candidate scheme set output by the large model, and 5, collecting user feedback and scheme execution data, and continuously optimizing the large model and a recommendation strategy.

Description

Technical Field

[0001] This invention discloses an intelligent recommendation method based on large-model fusion computing network services, which involves the intersection of artificial intelligence and large-model applications. Background Technology

[0002] With the booming development of the digital economy, the demand for the integration of cloud computing and network infrastructure is becoming increasingly urgent. As a core form of digital infrastructure, integrated computing and networking requires seamless collaboration between cloud resources such as servers, bare metal, VPCs, and public IP addresses, and network resources such as dedicated bandwidth, latency, and security protection, to support complex business scenarios such as high-definition video, industrial internet, and telemedicine. However, existing methods for recommending computing and networking services still have some shortcomings, such as:

[0003] Insufficient precision in demand parsing: User demands are often presented in the form of natural language, fuzzy descriptions, or business scenarios. Existing rule-based or traditional machine learning-based recommendation systems struggle to accurately extract key indicators, which can easily lead to discrepancies between demand and resource matching.

[0004] Resource status perception is lagging: Cloud resources are dynamic, and the status of cloud resource load, network bandwidth utilization, and leased line failure risk changes in real time. Existing methods rely on fixed-period collection or static rules, which cannot perceive resource dynamics in real time, resulting in poor timeliness of recommended solutions.

[0005] Weak multi-objective optimization capability: There is a conflict between user needs and the optimization objectives within the computing network. Traditional single-objective optimization or simple weighted methods are difficult to generate the globally optimal solution and cannot provide diverse candidate solutions for users to choose from.

[0006] Lack of dynamic adaptation mechanism: User business needs change with the scenario, such as the need for resource expansion during e-commerce promotion periods. Existing recommendation methods are difficult to dynamically adjust recommendation strategies based on user feedback and business changes, resulting in low reusability of the solution.

[0007] Insufficient integration of domain knowledge: Computing network services involve cloud technologies such as VPC architecture and load balancing strategies, and network technologies such as leased line QoS assurance and routing optimization. Existing models lack deep integration of domain knowledge, resulting in insufficient professionalism in recommended solutions. Summary of the Invention

[0008] This invention addresses the problems of existing technologies by providing an intelligent recommendation method based on a large-model fusion computing network service. Through domain adaptation of the large model, multimodal fusion reasoning, dynamic optimization, and feedback iteration mechanisms, it solves the pain points of existing computing network recommendation methods and achieves accurate matching and efficient recommendation of computing network resources.

[0009] The specific solution proposed in this invention is as follows:

[0010] This invention provides an intelligent recommendation method based on large-model fusion computing network services, comprising:

[0011] Step 1: Transform the user's raw input requirements into structured computing network requirement indicators.

[0012] Step 2: Collect and integrate the status information of cloud resources and network resources in real time to provide dynamic computing network resource data for large-scale model inference.

[0013] Step 3: Perform deep matching between user needs and computing network resources to generate a set of candidate solutions for multi-objective optimization:

[0014] Fine-tuning the large model for domain adaptation: The pre-trained large model is fine-tuned using a computational network domain corpus to optimize model parameters and enhance its understanding of computational network terminology; a computational network domain knowledge graph is constructed and injected into the large model; a computational network domain-specific prompt template is created to guide the large model to reason according to computational network logic.

[0015] User needs are transformed into user need vectors, and cross-modal matching is performed between these vectors and resource state vectors using the attention mechanism of a large model. This includes: semantic association matching (calculating the semantic similarity between need metrics and resource attributes); scenario adaptation matching (applying scenario knowledge from the large model for matching based on business scenarios); and constraint matching (selecting the most cost-effective resource combination based on the constraints in the reasoned needs).

[0016] Based on user needs, matching inference results, and internal network objectives, a Pareto optimal candidate set is generated using a large-scale model. This model learns target weights for different scenarios based on historical recommendation data and user feedback. When objectives conflict, the model selects a compromise recommendation. Finally, the model generates 3-5 differentiated candidate solutions, each including cloud resource configuration, network resource configuration, security policies, and cost details.

[0017] Step 4: Optimize the candidate solution set output by the large model:

[0018] Based on the rule base and historical cases in the field of computing networks, the technical feasibility of the candidate solutions is verified.

[0019] Quantitative analysis of candidate solutions' metrics: performance metrics analysis, including predicted leased line latency, cloud server response speed, and overall system availability; cost metrics analysis, including calculation of average monthly cost and cost breakdown; and security metrics analysis.

[0020] Candidate solutions are sorted according to user's historical preferences;

[0021] Step 5: Collect user feedback and implementation data to continuously optimize the large model and recommendation strategy.

[0022] Furthermore, step 1 of the intelligent recommendation method based on large-model fusion computing network services involves transforming the user's original input requirements into structured computing network requirement indicators.

[0023] It receives user requests in various input formats forwarded by the UI interaction layer, including natural language text, semi-structured forms, and scene template selection results.

[0024] Based on a pre-trained large model, user needs are analyzed: key entities are extracted from user needs, and implicit user needs are inferred through the context modeling function of the large model. The analyzed needs are then mapped to the standard indicator system of the computing network domain to obtain computing network demand indicators.

[0025] Based on the knowledge base of computing networks, the rationality of the parsed computing network requirement indicators is verified.

[0026] Furthermore, step 2 of the intelligent recommendation method based on large-model fusion computing network services specifically includes:

[0027] Connect to cloud resource platforms and network resource management platforms to collect multi-dimensional resource data: acquire cloud resource data, network resource data, and environmental data.

[0028] By employing the multimodal fusion capabilities of a large model, structured numerical data, unstructured text data, and topological data are transformed into unified resource state vectors. Through the embedding layer of the large model, different types of resource state vectors are mapped to the same semantic space, enabling cross-modal resource state data association and fusion.

[0029] Based on the fused resource status data, the classification and regression functions of the large model are used to assess the current status level of the resources, and a comprehensive status score is output through large model inference.

[0030] Based on the network demand indicators, a set of candidate resources that meet the basic conditions is selected.

[0031] Furthermore, in step 2 of the intelligent recommendation method based on large-model fusion computing network services, the following formula is used in the embedding layer of the large model:

[0032] U = LM - Embedding(T; ΘLM)

[0033] Different types of resource state vectors are mapped to the same semantic space, where T is the original requirement text input by the user;

[0034] ΘLM represents the large model parameters after domain fine-tuning, which incorporates knowledge of the computational network domain.

[0035] LM-Embedding(·) is the text embedding function for large models, which outputs a semantic vector of requirements U∈Rd, where d is the vector dimension, usually 768 or 1024. The vector dimension includes business type, core indicators, and implicit requirements.

[0036] Furthermore, in step 2 of the intelligent recommendation method based on large-model fusion computing network services, the following formula is used:

[0037] R=α·Num-Embedding(N)+(1-α)·Text-Embedding(L; ΘLM)

[0038] Cross-modal resource state data association and fusion are performed, where N is structured resource numerical data; L is unstructured resource text data; Num-Embedding(·) is a numerical embedding function that converts numerical values ​​into vectors; Text-Embedding(·) is a large model text embedding function that processes text data; α∈[0,1] is the fusion weight; the output resource state vector R∈Rd is consistent with the dimension of the demand vector U, which facilitates subsequent matching.

[0039] Furthermore, in step 3 of the intelligent recommendation method based on large-model fusion computing network services, semantic association matching is performed, including:

[0040] Using the formula:

[0041] Sim(U,R)=∥U∥·∥R∥U·Attention(R;U)

[0042] Calculate the semantic similarity between demand indicators and resource attributes, where Attention(R;U) is the attention function of the large model, dynamically adjusting the feature weights of the resource vector R based on the demand vector U;

[0043] · represents the vector inner product, and ∥·∥ represents the L2 norm;

[0044] The output matching degree Sim∈[0,1], the higher the value, the better the resource matches the demand.

[0045] Furthermore, in step 3 of the intelligent recommendation method based on large-model fusion computing network services, generating the Pareto optimal candidate set through the large model includes:

[0046] Create the objective function:

[0047] minS[λ1·C(S)+λ2·(1-Q(S))+λ3·(1-S(S))]

[0048] Where S represents the candidate scheme;

[0049] C(S) is the cost of the solution, normalized to [0,1]);

[0050] Q(S) is the performance index of the scheme;

[0051] S(S) represents the security level of the scheme, normalized to [0,1]);

[0052] λ1, λ2, λ3 are the target weights;

[0053] The objective function is solved using a large model via gradient descent or evolutionary algorithms, generating 3-5 sets of Pareto optimal solutions {S1,S2,...,Sk}.

[0054] Furthermore, step 5 of the intelligent recommendation method based on large-model fusion computing network services specifically includes:

[0055] Collect both explicit and implicit user feedback on the recommended solutions, and also collect operational data after the solutions are implemented;

[0056] By analyzing feedback data using large-scale models, we can pinpoint recommendation problems.

[0057] If user feedback indicates that costs exceed expectations, analyze whether the cost budget extraction in the requirements analysis is accurate and whether a high-cost-performance combination was selected in resource matching.

[0058] If the latency does not meet the target after the solution is implemented, analyze whether the resource status perception is lagging and whether the network load is underestimated during inference.

[0059] Based on the feedback analysis results, incremental optimization is performed on the large model:

[0060] To address discrepancies in the requirement analysis, supplementary labeled data for the corresponding scenarios was added for fine-tuning.

[0061] To address resource matching discrepancies, update the domain knowledge graph and prompt template.

[0062] To address the bias in multi-objective weights, the target weight learning mechanism is adjusted.

[0063] The recommendation strategy will be updated based on the feedback and optimization results.

[0064] Furthermore, in step 5 of the intelligent recommendation method based on large-model fusion computing network services, the parameters of the large model are optimized based on user feedback data, using the following formula:

[0065]

[0066] Where Θold represents the model parameters before optimization;

[0067] L(U,S,F) is the feedback loss function, where the actual deviation between user demand U and recommended solution S is calculated from the feedback data F.

[0068] This represents the gradient of the loss function with respect to the parameters.

[0069] η is the learning rate;

[0070] Output the optimized model parameters Θnew to improve the accuracy of the next round of recommendations.

[0071] This invention also provides an intelligent recommendation system based on large-model fusion computing network services, including a user demand analysis module, a computing network resource perception module, a large-model fusion reasoning module, a recommendation scheme generation and optimization module, and a feedback iteration module.

[0072] The user requirement parsing module transforms the raw requirements input by the user into structured computing network requirement indicators:

[0073] The computing network resource awareness module collects and integrates real-time status information of cloud and network resources, providing dynamic computing network resource data for large-scale model inference.

[0074] The large-scale model fusion inference module performs deep matching between user needs and computing network resources, generating a set of candidate solutions for multi-objective optimization:

[0075] Fine-tuning the large model for domain adaptation: The pre-trained large model is fine-tuned using a computational network domain corpus to optimize model parameters and enhance its understanding of computational network terminology; a computational network domain knowledge graph is constructed and injected into the large model; a computational network domain-specific prompt template is created to guide the large model to reason according to computational network logic.

[0076] User needs are transformed into user need vectors, and cross-modal matching is performed between these vectors and resource state vectors using the attention mechanism of a large model. This includes: semantic association matching (calculating the semantic similarity between need metrics and resource attributes); scenario adaptation matching (applying scenario knowledge from the large model for matching based on business scenarios); and constraint matching (selecting the most cost-effective resource combination based on the constraints in the reasoned needs).

[0077] Based on user needs, matching inference results, and internal network objectives, a Pareto optimal candidate set is generated using a large-scale model. This model learns target weights for different scenarios based on historical recommendation data and user feedback. When objectives conflict, the model selects a compromise recommendation. Finally, the model generates 3-5 differentiated candidate solutions, each including cloud resource configuration, network resource configuration, security policies, and cost details.

[0078] The recommendation generation and optimization module optimizes the candidate solution set output by the large model:

[0079] Based on the rule base and historical cases in the field of computing networks, the technical feasibility of the candidate solutions is verified.

[0080] Quantitative analysis of candidate solutions' metrics: performance metrics analysis, including predicted leased line latency, cloud server response speed, and overall system availability; cost metrics analysis, including calculation of average monthly cost and cost breakdown; and security metrics analysis.

[0081] Candidate solutions are sorted according to user's historical preferences;

[0082] The feedback iteration module collects user feedback and solution execution data to continuously optimize the large model and recommendation strategy.

[0083] Compared with the prior art, the present invention achieves the following technical effects through the above technical solution:

[0084] Improved accuracy of requirement parsing: The semantic understanding and intent reasoning capabilities of the large model enable the accuracy of requirement parsing to be ≥95% and the coverage of implicit requirement mining to be ≥90%, solving the problem of insufficient handling of fuzzy requirements by existing methods.

[0085] Enhanced resource matching timeliness: Multimodal fusion perception and dynamic adjustment reasoning reduce the resource status perception delay to ≤5 seconds and the dynamic adjustment response time to ≤100ms, adapting to the dynamic changes of computing network resources.

[0086] Significant multi-objective optimization results: The generated Pareto optimal solution set can simultaneously satisfy user and internal network objectives, improve resource utilization by ≥20%, reduce user costs by ≥15%, and achieve 100% security compliance.

[0087] User experience optimization: Personalized sorting and visual display improve user solution selection efficiency by ≥30%; the requirement feedback and optimization suggestion function reduces the unreasonable requirement rate by ≥40%.

[0088] Enhanced scalability: Domain adaptation and design of large models support rapid integration of new business scenarios, with a scenario expansion cycle of ≤7 days. Attached Figure Description

[0089] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0090] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0091] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0092] Example 1

[0093] This invention provides an intelligent recommendation method based on large-model fusion computing network services, comprising:

[0094] Step 1: Transform the user's raw input requirements into structured computational network requirement indicators, providing a foundation for subsequent matching. This relies heavily on the natural language understanding and domain knowledge reasoning capabilities of the large-scale model, specifically including:

[0095] Receive user requirements in various input formats forwarded by the UI interaction layer, including natural language text such as "Build a cloud office system for a company with 50 employees, with secure data transmission and a monthly budget of no more than 20,000 yuan"; semi-structured forms such as those containing fields for business type, user size, and region; and scenario template selection results.

[0096] Based on pre-trained large models such as LLaMA and GPT series, user needs are analyzed: key entities in user needs are extracted, including business type such as office, live broadcast, industrial control, user scale such as 50-person enterprise, large park, geographical range such as East China region, cross-regional, core indicators such as latency <20ms, bandwidth ≥100Mbps.

[0097] By leveraging the contextual modeling capabilities of large models, we can infer implicit user needs. For example, we can infer the requirements for high bandwidth, low jitter, and data encryption from medical image transmission scenarios; and we can infer the requirements for multi-regional node deployment and cross-border dedicated line optimization from cross-border e-commerce platforms.

[0098] The parsed requirements are mapped to the standard indicator system in the computing network field to obtain computing network requirement indicators, such as converting high transmission speed into dedicated line bandwidth ≥500Mbps and one-way latency ≤30ms; data security into support for AES-256 encryption, VPC isolation, and WAF protection.

[0099] Based on a knowledge base in the field of computing networks, the rationality of the parsed computing network requirement indicators is verified. For example, if a user proposes "deploying ultra-low latency leased lines (<10ms) in remote areas", combined with geographical network topology knowledge, it is determined that this requirement is difficult to achieve under the current network architecture, and it is marked as a "requirement to be optimized", triggering the subsequent feedback process.

[0100] Step 2: Real-time collection and integration of cloud and network resource status information to provide dynamic computing network resource data for large-scale model inference, specifically including:

[0101] Connect to cloud resource platforms and network resource management platforms to collect multi-dimensional resource data:

[0102] Obtain cloud resource data: cloud hosts such as CPU utilization, memory usage, and disk I / O; bare metal servers such as hardware configuration and load status; VPC such as network segment planning and routing tables; public IPs such as availability and bandwidth limit; load balancers such as concurrent connections and forwarding efficiency; and cloud disks such as capacity and read / write speed.

[0103] Network resource data: leased line resources such as bandwidth utilization, current latency, packet loss rate, QoS level; network topology such as node connection relationships and routing paths; fault records such as historical outage duration and fault causes; security status such as firewall rules and intrusion detection logs.

[0104] Environmental data includes geographical information such as data center location and network coverage, time information such as peak hours and holidays, and policy information such as bandwidth tariff adjustments and security compliance requirements.

[0105] Employing the multimodal fusion capabilities of a large model, structured numerical data, unstructured text data, and topological data are transformed into unified resource state vectors. Through the embedding layer of the large model, these different types of resource state vectors are mapped to the same semantic space, enabling cross-modal resource state data association and fusion. The embedding layer of the large model utilizes the following formula:

[0106] U = LM - Embedding(T; ΘLM)

[0107] Different types of resource state vectors are mapped to the same semantic space, where T is the original requirement text input by the user;

[0108] ΘLM represents the large model parameters after domain fine-tuning, which incorporates knowledge of the computational network domain.

[0109] LM-Embedding(·) is the text embedding function for large models, which outputs a semantic vector of requirements U∈Rd, where d is the vector dimension, usually 768 or 1024. The vector dimension includes business type, core indicators, and implicit requirements.

[0110] Using the formula:

[0111] R=α·Num-Embedding(N)+(1-α)·Text-Embedding(L; ΘLM)

[0112] Cross-modal resource state data association and fusion are performed, where N is structured resource numerical data; L is unstructured resource text data; Num-Embedding(·) is a numerical embedding function that converts numerical values ​​into vectors; Text-Embedding(·) is a large model text embedding function that processes text data; α∈[0,1] is the fusion weight; the output resource state vector R∈Rd is consistent with the dimension of the demand vector U, which facilitates subsequent matching.

[0113] Based on the fused resource status data, the classification and regression functions of the large model are used to assess the current status level of resources, such as excellent, good, medium, and poor. For example, for dedicated line resources, combined with indicators such as bandwidth utilization <60% as excellent, latency <20ms as excellent, and packet loss rate <0.1% as excellent, the large model infers and outputs a comprehensive status score. For cloud hosts, based on the load trend prediction of CPU, memory, and disk load for the next hour, resources with "high load risk" are marked.

[0114] Based on the network demand indicators, a set of candidate resources that meet the basic conditions is selected. For example, if the user's requirement is "latency < 30ms", then dedicated line resources with current latency > 30ms are filtered out; if the requirement is "supports GPU acceleration", then cloud host resources configured with GPU are selected.

[0115] Step 3: Perform deep matching between user needs and computing network resources to generate a set of candidate solutions for multi-objective optimization:

[0116] Fine-tuning the large model for domain adaptation: The pre-trained large model is fine-tuned using a corpus of computer science domain terminology to optimize model parameters and enhance its understanding of computer science terminology; a knowledge graph of the computer science domain is constructed and injected into the large model; a dedicated prompt template for the computer science domain is created to guide the large model in reasoning according to computer science logic. For example, a prompt for demand-based reasoning would be: "User demand: {demand text}. Please parse out the business type, core metrics (bandwidth, latency, security level, cost), and implicit requirements in JSON format."

[0117] User requirements are transformed into user requirement vectors, and cross-modal matching is performed between these vectors and resource state vectors using the attention mechanism of a large model. Semantic association matching is then performed: the semantic similarity between requirement metrics and resource attributes is calculated, for example, associating data security transmission requirements with dedicated lines supporting encrypted transmission and VPC-isolated cloud hosts. Semantic association matching includes:

[0118] Using the formula:

[0119] Sim(U,R)=∥U∥·∥R∥U·Attention(R;U)

[0120] Calculate the semantic similarity between demand indicators and resource attributes, where Attention(R;U) is the attention function of the large model, dynamically adjusting the feature weights of the resource vector R based on the demand vector U;

[0121] · represents the vector inner product, and ∥·∥ represents the L2 norm;

[0122] The output matching degree Sim∈[0,1], the higher the value, the better the resource matches the demand.

[0123] Perform scenario adaptation and matching: For business scenarios, call the scenario knowledge of the large model for matching. For example, the industrial Internet scenario needs to match low latency dedicated line (<10ms) + edge computing node + high reliability cloud host (availability 99.99%).

[0124] Perform constraint matching: Based on the constraints in the reasoning requirements, select the resource combination with the best cost performance. For example, under the constraint of cost ≤ 20,000 yuan / month and bandwidth ≥ 500Mbps, select the resource combination with the best cost performance.

[0125] Based on user needs, matching inference results, and internal network objectives, a Pareto optimal candidate set is generated through a large model: the large model learns the target weights under different scenarios based on historical recommendation data and user feedback; when objectives conflict, the large model selects a compromise solution; and the large model generates 3-5 sets of differentiated candidate solutions, each set including cloud resource configuration, network resource configuration, security strategy, and cost details.

[0126] Generate a Pareto optimal candidate set using a large model, including:

[0127] Create the objective function:

[0128] minS[λ1·C(S)+λ2·(1-Q(S))+λ3·(1-S(S))]

[0129] Where S represents the candidate scheme;

[0130] C(S) is the cost of the solution, normalized to [0,1]);

[0131] Q(S) is the performance index of the scheme;

[0132] S(S) represents the security level of the scheme, normalized to [0,1]);

[0133] λ1, λ2, λ3 are the target weights;

[0134] The objective function is solved using a large model via gradient descent or evolutionary algorithms, generating 3-5 sets of Pareto optimal solutions {S1,S2,...,Sk}.

[0135] The recommendation scheme is dynamically adjusted based on real-time changes in resource status and updates in user needs.

[0136] Resource fluctuation adaptation: If the recommended dedicated line resources suddenly encounter excessive load, the real-time inference of the large model will replace the solution, such as switching to a backup dedicated line in the same area;

[0137] Adapting to Changes in Requirements: If users temporarily add disaster recovery and backup requirements, the large model will be quickly updated to include cross-regional cloud disk backup and dedicated line redundant link configuration.

[0138] Step 4: Optimize the candidate solution set output by the large model:

[0139] Based on the rule base and historical cases in the computing network field, the technical feasibility of candidate solutions is verified. For example, the matching between cloud server specifications and leased line bandwidth is checked to avoid performance bottlenecks caused by high-bandwidth leased lines and low-configuration cloud servers. The rationality of the network topology is also verified, such as whether VPC network segments conflict and whether the routing configuration is reachable.

[0140] Quantitative analysis of candidate solutions' metrics: performance metrics analysis, including predicted leased line latency, cloud server response speed, and overall system availability; cost metrics analysis, including calculation of average monthly cost and cost breakdown; and security metrics analysis.

[0141] Candidate solutions are sorted based on users' historical preferences. For example, price-sensitive users are sorted by cost from low to high, and performance-sensitive users are sorted by latency from low to high.

[0142] Step 5: Collect user feedback and implementation data to continuously optimize the large model and recommendation strategy. This includes:

[0143] Collect both explicit and implicit user feedback on the recommended solutions, and also collect operational data after the solutions are implemented;

[0144] By analyzing feedback data using large-scale models, we can pinpoint recommendation problems.

[0145] If user feedback indicates that costs exceed expectations, analyze whether the cost budget extraction in the requirements analysis is accurate and whether a high-cost-performance combination was selected in resource matching.

[0146] If the latency does not meet the target after the solution is implemented, analyze whether the resource status perception is lagging and whether the network load is underestimated during inference.

[0147] Based on the feedback analysis results, incremental optimization is performed on the large model:

[0148] To address discrepancies in the requirement analysis, supplementary labeled data for the corresponding scenarios was added for fine-tuning.

[0149] To address resource matching discrepancies, update the domain knowledge graph and prompt template.

[0150] To address the bias in multi-objective weights, the target weight learning mechanism is adjusted.

[0151] Based on the feedback and optimization results, the recommendation strategy is updated, such as adjusting resource filtering thresholds, optimizing the solution ranking algorithm, and adding new scenario adaptation rules. The optimization of large model parameters based on user feedback data is illustrated by the following formula:

[0152]

[0153] Where Θold represents the model parameters before optimization;

[0154] L(U,S,F) is the feedback loss function, where the actual deviation between user demand U and recommended solution S is calculated from the feedback data F.

[0155] This represents the gradient of the loss function with respect to the parameters.

[0156] η is the learning rate;

[0157] Output the optimized model parameters Θnew to improve the accuracy of the next round of recommendations.

[0158] Example 2

[0159] This invention also provides an intelligent recommendation system based on large-model fusion computing network services, including a user demand analysis module, a computing network resource perception module, a large-model fusion reasoning module, a recommendation scheme generation and optimization module, and a feedback iteration module.

[0160] The user requirement parsing module transforms the raw requirements input by the user into structured computing network requirement indicators:

[0161] The computing network resource awareness module collects and integrates real-time status information of cloud and network resources, providing dynamic computing network resource data for large-scale model inference.

[0162] The large-scale model fusion inference module performs deep matching between user needs and computing network resources, generating a set of candidate solutions for multi-objective optimization:

[0163] Fine-tuning the large model for domain adaptation: The pre-trained large model is fine-tuned using a computational network domain corpus to optimize model parameters and enhance its understanding of computational network terminology; a computational network domain knowledge graph is constructed and injected into the large model; a computational network domain-specific prompt template is created to guide the large model to reason according to computational network logic.

[0164] User needs are transformed into user need vectors, and cross-modal matching is performed between these vectors and resource state vectors using the attention mechanism of a large model. This includes: semantic association matching (calculating the semantic similarity between need metrics and resource attributes); scenario adaptation matching (applying scenario knowledge from the large model for matching based on business scenarios); and constraint matching (selecting the most cost-effective resource combination based on the constraints in the reasoned needs).

[0165] Based on user needs, matching inference results, and internal network objectives, a Pareto optimal candidate set is generated using a large-scale model. This model learns target weights for different scenarios based on historical recommendation data and user feedback. When objectives conflict, the model selects a compromise recommendation. Finally, the model generates 3-5 differentiated candidate solutions, each including cloud resource configuration, network resource configuration, security policies, and cost details.

[0166] The recommendation generation and optimization module optimizes the candidate solution set output by the large model:

[0167] Based on the rule base and historical cases in the field of computing networks, the technical feasibility of the candidate solutions is verified.

[0168] Quantitative analysis of candidate solutions' metrics: performance metrics analysis, including predicted leased line latency, cloud server response speed, and overall system availability; cost metrics analysis, including calculation of average monthly cost and cost breakdown; and security metrics analysis.

[0169] Candidate solutions are sorted according to user's historical preferences;

[0170] The feedback iteration module collects user feedback and solution execution data to continuously optimize the large model and recommendation strategy.

[0171] This invention's system can be applied to a layered architecture, which includes a UI interaction layer, a business logic layer, a data support layer, and a resource interface layer. Each layer collaborates through standardized interfaces to achieve intelligent processing from user request input to recommended solution output. Specifically, this invention's system can be deployed in the business logic layer, which is the core layer, integrating a large-model fusion inference module and leading the processes of requirement parsing, resource matching, solution generation, and optimization iteration.

[0172] The information interaction and execution process between the modules in the above system are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description in the method embodiment of the present invention, and will not be repeated here.

[0173] Similarly, the system of the present invention achieves the following technical effects through the above technical solution:

[0174] Improved accuracy of requirement parsing: The semantic understanding and intent reasoning capabilities of the large model enable the accuracy of requirement parsing to be ≥95% and the coverage of implicit requirement mining to be ≥90%, solving the problem of insufficient handling of fuzzy requirements by existing methods.

[0175] Enhanced resource matching timeliness: Multimodal fusion perception and dynamic adjustment reasoning reduce the resource status perception delay to ≤5 seconds and the dynamic adjustment response time to ≤100ms, adapting to the dynamic changes of computing network resources.

[0176] Significant multi-objective optimization results: The generated Pareto optimal solution set can simultaneously satisfy user and internal network objectives, improve resource utilization by ≥20%, reduce user costs by ≥15%, and achieve 100% security compliance.

[0177] User experience optimization: Personalized sorting and visual display improve user solution selection efficiency by ≥30%; the requirement feedback and optimization suggestion function reduces the unreasonable requirement rate by ≥40%.

[0178] Enhanced scalability: Domain adaptation and design of large models support rapid integration of new business scenarios, with a scenario expansion cycle of ≤7 days.

[0179] It should be noted that not all steps and modules in the above processes and system structures are mandatory; some steps or modules can be omitted as needed. The execution order of the steps is not fixed and can be adjusted as required. The system structures described in the above embodiments can be physical or logical structures. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be implemented by certain components in multiple independent devices.

[0180] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. An intelligent recommendation method based on large-scale model fusion computing network services, characterized by: include: Step 1: Transform the user's raw input requirements into structured computing network requirement indicators. Step 2: Collect and integrate the status information of cloud resources and network resources in real time to provide dynamic computing network resource data for large-scale model inference. Step 3: Perform deep matching between user needs and computing network resources to generate a set of candidate solutions for multi-objective optimization: Fine-tuning the large model for domain adaptation: The pre-trained large model is fine-tuned using a computational network domain corpus to optimize model parameters and enhance its understanding of computational network terminology; a computational network domain knowledge graph is constructed and injected into the large model; a computational network domain-specific prompt template is created to guide the large model to reason according to computational network logic. User needs are transformed into user need vectors, and cross-modal matching is performed between these vectors and resource state vectors using the attention mechanism of a large model. This includes: semantic association matching (calculating the semantic similarity between need metrics and resource attributes); scenario adaptation matching (applying scenario knowledge from the large model for matching based on business scenarios); and constraint matching (selecting the most cost-effective resource combination based on the constraints in the reasoned needs). Based on user needs, matching inference results, and internal network objectives, a Pareto optimal candidate set is generated using a large-scale model. This model learns target weights for different scenarios based on historical recommendation data and user feedback. When objectives conflict, the model selects a compromise recommendation. Finally, the model generates 3-5 differentiated candidate solutions, each including cloud resource configuration, network resource configuration, security policies, and cost details. Step 4: Optimize the candidate solution set output by the large model: Based on the rule base and historical cases in the field of computing networks, the technical feasibility of the candidate solutions is verified. Quantitative analysis of candidate solutions' metrics: performance metrics analysis, including predicted leased line latency, cloud server response speed, and overall system availability; cost metrics analysis, including calculation of average monthly cost and cost breakdown; and security metrics analysis. Candidate solutions are sorted according to user's historical preferences; Step 5: Collect user feedback and implementation data to continuously optimize the large model and recommendation strategy.

2. The intelligent recommendation method based on large-scale model fusion computing network services according to claim 1, characterized in that step 1: the user's original input requirements are transformed into structured computing network requirement indicators: It receives user requests in various input formats forwarded by the UI interaction layer, including natural language text, semi-structured forms, and scene template selection results. Based on a pre-trained large model, user needs are analyzed: key entities are extracted from user needs, and implicit user needs are inferred through the context modeling function of the large model. The analyzed needs are then mapped to the standard indicator system of the computing network domain to obtain computing network demand indicators. Based on the knowledge base of computing networks, the rationality of the parsed computing network requirement indicators is verified.

3. The intelligent recommendation method based on large-model fusion computing network services according to claim 1, characterized in that: Step 2 specifically includes: Connect to cloud resource platforms and network resource management platforms to collect multi-dimensional resource data: acquire cloud resource data, network resource data, and environmental data. By employing the multimodal fusion capabilities of a large model, structured numerical data, unstructured text data, and topological data are transformed into unified resource state vectors. Through the embedding layer of the large model, different types of resource state vectors are mapped to the same semantic space, enabling cross-modal resource state data association and fusion. Based on the fused resource status data, the classification and regression functions of the large model are used to assess the current status level of the resources, and a comprehensive status score is output through large model inference. Based on the network demand indicators, a set of candidate resources that meet the basic conditions is selected.

4. The intelligent recommendation method based on large-model fusion computing network services according to claim 3, characterized in that in step 2, the following formula is used in the embedding layer of the large model: U = LM - Embedding(T; ΘLM) Different types of resource state vectors are mapped to the same semantic space, where... T represents the original request text entered by the user; ΘLM represents the large model parameters after domain fine-tuning, which incorporates knowledge of the computational network domain. LM-Embedding(·) is the text embedding function for large models, which outputs a semantic vector of requirements U∈Rd, where d is the vector dimension, usually 768 or 1024. The vector dimension includes business type, core indicators, and implicit requirements.

5. The intelligent recommendation method based on large-model fusion computing network services according to claim 3 or 4, characterized in that: Step 2 uses the following formula: R=α·Num-Embedding(N)+(1-α)·Text-Embedding(L; ΘLM) Cross-modal resource state data association and fusion are performed, where N is structured resource numerical data; L is unstructured resource text data; Num-Embedding(·) is a numerical embedding function that converts numerical values ​​into vectors; Text-Embedding(·) is a large model text embedding function that processes text data; α∈[0,1] is the fusion weight; the output resource state vector R∈Rd is consistent with the dimension of the demand vector U, which facilitates subsequent matching.

6. The intelligent recommendation method based on large-model fusion computing network services according to claim 1, characterized in that: Step 3 involves semantic association matching, including: Using the formula: Sim(U,R)=∥U∥·∥R∥U·Attention(R;U) Calculate the semantic similarity between demand indicators and resource attributes, where Attention(R;U) is the attention function of the large model, dynamically adjusting the feature weights of the resource vector R based on the demand vector U; · represents the vector inner product, and ∥·∥ represents the L2 norm; The output matching degree Sim∈[0,1], the higher the value, the better the resource matches the demand.

7. The intelligent recommendation method based on large-model fusion computing network services according to claim 1, characterized in that: Step 3 generates the Pareto optimal candidate set using a large model, including: Create the objective function: minS[λ1·C(S)+λ2·(1-Q(S))+λ3·(1-S(S))] Where S represents the candidate scheme; C(S) is the cost of the solution, normalized to [0,1]); Q(S) is the performance index of the scheme; S(S) represents the security level of the scheme, normalized to [0,1]); λ1, λ2, λ3 are the target weights; The objective function is solved using a large model via gradient descent or evolutionary algorithms, generating 3-5 sets of Pareto optimal solutions {S1,S2,...,Sk}.

8. The intelligent recommendation method based on large-model fusion computing network services according to claim 1, characterized in that: Step 5 specifically includes: Collect both explicit and implicit user feedback on the recommended solutions, and also collect operational data after the solutions are implemented; By analyzing feedback data using large-scale models, we can pinpoint recommendation problems. If user feedback indicates that costs exceed expectations, analyze whether the cost budget extraction in the requirements analysis is accurate and whether a high-cost-performance combination was selected in resource matching. If the latency does not meet the target after the solution is implemented, analyze whether the resource status perception is lagging and whether the network load is underestimated during inference. Based on the feedback analysis results, incremental optimization is performed on the large model: To address discrepancies in the requirement analysis, supplementary labeled data for the corresponding scenarios was added for fine-tuning. To address resource matching discrepancies, update the domain knowledge graph and prompt template. To address the bias in multi-objective weights, the target weight learning mechanism is adjusted. The recommendation strategy will be updated based on the feedback and optimization results.

9. The intelligent recommendation method based on large-model fusion computing network services according to claim 1, characterized in that: In step 5, the parameters of the large model are optimized based on user feedback data, using the following formula: Where Θold represents the model parameters before optimization; L(U,S,F) is the feedback loss function, where the actual deviation between user demand U and recommended solution S is calculated from the feedback data F. This represents the gradient of the loss function with respect to the parameters. η is the learning rate; Output the optimized model parameters Θnew to improve the accuracy of the next round of recommendations.

10. An intelligent recommendation system based on large-scale model fusion computing network services, characterized by: It includes a user needs analysis module, a computing network resource awareness module, a large model fusion and inference module, a recommendation scheme generation and optimization module, and a feedback iteration module. The user requirement analysis module transforms the raw requirements input by users into structured computing network requirement indicators. The computing network resource awareness module collects and integrates real-time status information of cloud and network resources, providing dynamic computing network resource data for large-scale model inference. The large-scale model fusion inference module performs deep matching between user needs and computing network resources, generating a set of candidate solutions for multi-objective optimization: Fine-tuning the large model for domain adaptation: The pre-trained large model is fine-tuned using a computational network domain corpus to optimize model parameters and enhance its understanding of computational network terminology; a computational network domain knowledge graph is constructed and injected into the large model; a computational network domain-specific prompt template is created to guide the large model to reason according to computational network logic. User needs are transformed into user need vectors, and cross-modal matching is performed between these vectors and resource state vectors using the attention mechanism of a large model. This includes: semantic association matching (calculating the semantic similarity between need metrics and resource attributes); scenario adaptation matching (applying scenario knowledge from the large model for matching based on business scenarios); and constraint matching (selecting the most cost-effective resource combination based on the constraints in the reasoned needs). Based on user needs, matching inference results, and internal network objectives, a Pareto optimal candidate set is generated using a large-scale model. This model learns target weights for different scenarios based on historical recommendation data and user feedback. When objectives conflict, the model selects a compromise recommendation. Finally, the model generates 3-5 differentiated candidate solutions, each including cloud resource configuration, network resource configuration, security policies, and cost details. The recommendation generation and optimization module optimizes the candidate solution set output by the large model: Based on the rule base and historical cases in the field of computing networks, the technical feasibility of the candidate solutions is verified. Quantitative analysis of candidate solutions' metrics: performance metrics analysis, including predicted leased line latency, cloud server response speed, and overall system availability; cost metrics analysis, including calculation of average monthly cost and cost breakdown; and security metrics analysis. Candidate solutions are sorted according to user's historical preferences; The feedback iteration module collects user feedback and solution execution data to continuously optimize the large model and recommendation strategy.

Images