An information technology service intelligent matching method based on a knowledge graph

Abstract

The application discloses a kind of information technology service intelligent matching methods based on knowledge graph, solve the defect of traditional matching in low-carbon scheduling, multimodal demand and dynamic response.By top-down and bottom-up construction ontology, encode multi-source data using graph neural network, use static ZIP model and dynamic induction motor model to depict service capability, filter core entities through multi-head attention mechanism;Improved SHAP model and cross-modal attention mechanism are used to analyze multi-modal data, construct evolution graph and generate demand vector;Based on cosine similarity matching, through double DQN network adaptive adjustment strategy, Adam optimizer generates multi-objective scheme and verifies robustness;Use kernel function to estimate difference, update service features, adjust entity relationship, expand ontology and optimize algorithm, form a closed loop.The application realizes low-carbon configuration, multimodal analysis and strategy optimization, improves matching accuracy, robustness and evolution ability.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent matching of information technology services and knowledge graph application, and specifically relates to an information technology service intelligent matching method based on knowledge graph driving. BACKGROUND

[0002] In the field of information technology services, traditional matching methods mainly rely on static rules or simple keyword matching, which are difficult to cope with the dynamic complexity of service resources and user needs. The specific defects are reflected in the following three aspects:

[0003] Technical gap of low-carbon scheduling and resource optimization: the existing technology does not deeply integrate the low-carbon goal into the service matching process, and lacks quantitative modeling of resource consumption and carbon emissions in service operation. For example, when scheduling services, the traditional scheme only focuses on function matching and ignores the resource use efficiency of service instances, resulting in high energy consumption services being frequently called, which is difficult to meet the policy requirements of low-carbon operation of distribution networks. This "emphasis on function and neglect of energy efficiency" mode makes information technology services in large-scale applications likely to exacerbate energy waste, which is out of line with the development trend of green computing.

[0004] Shallow and lagging interaction mechanism of source-load-storage: the linkage between user needs, service resources and service capacity reserves stays at the static mapping level. On the one hand, user needs are mainly single modal, lacking integrated analysis of multi-form needs such as images and voices, resulting in one-sided understanding of needs; on the other hand, the dynamic performance data of service resources are not effectively included in the matching model, which cannot reflect the "reserve" state of service capacity in real time. For example, when the CPU usage rate exceeds 90% during the service peak period, the traditional algorithm cannot adjust the matching strategy in time, resulting in a decline in service quality and damage to user satisfaction.

[0005] Double bottleneck of algorithm efficiency and dynamic response: traditional matching algorithms are mostly based on static similarity calculation or heuristic rules, which are difficult to cope with the rapid changes of multi-modal needs and the spatio-temporal evolution of service characteristics. The specific performance is as follows:

[0006] Single feature representation: only relying on basic attributes of services for matching, ignoring the dynamic evolution characteristics of service capacity;

[0007] Lag in strategy adjustment: lack of adaptive mechanism, unable to optimize the matching strategy according to real-time feedback. For example, when user needs are upgraded from "basic data processing" to "real-time data analysis", the traditional algorithm needs to be manually modified to respond, resulting in a delay in matching.

[0008] Therefore, an information technology service intelligent matching method based on knowledge graph driving is proposed. SUMMARY

[0009] According to the deficiencies in the prior art above, the purpose of the present application is to solve the problems of static, single mode and low efficiency of traditional methods by constructing a dynamic knowledge graph, multi-modal demand analysis, intelligent matching optimization and closed-loop iteration mechanism.

[0010] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0011] A knowledge graph driven information technology service intelligent matching method, comprising the following steps:

[0012] S1. Multi-source heterogeneous data knowledge graph construction, combining top-down and bottom-up methods to construct the ontology of the information technology service field, collecting multi-source data and encoding them into an initial graph by a graph neural network;

[0013] A parameter identification method based on a static ZIP model and a dynamic induction motor model is used to establish the static characteristics and dynamic evolution characteristics of service capabilities, filter high correlation entities by a multi-head attention mechanism, form a service network, and provide a resource base for subsequent steps;

[0014] S2. Multi-modal demand feature dynamic extraction, based on the initial graph formed in S1, using an improved SHAP model and a cross-modal attention mechanism to extract multi-modal features of user demand, constructing an evolution graph, identifying demand change nodes by threshold comparison, and generating demand vectors as key matching basis;

[0015] S3. Service and demand dynamic matching optimization, calculating the similarity between the demand vector in S2 and the initial graph features in step one, using a double DQN network to adaptively adjust the matching strategy, using an Adam optimizer to generate a target function, and forming a multi-objective matching scheme through adversarial sample verification;

[0016] S4. Matching result driven graph iteration, based on the matching results in S3 and user feedback, estimating the difference using a kernel function, updating service features based on dynamic load modeling, improving the matching algorithm through an optimization closed-loop mechanism, and using the optimized graph for new matching to form a cycle.

[0017] Further, in S1, the specific steps of combining top-down and bottom-up methods to construct the ontology of the information technology service field are as follows:

[0018] S111. Top-down design, organizing field experts and technical personnel, based on information technology service industry standards and business logic, combing out a basic concept framework, defining "information technology service" as the top class, with "software development service", "system operation and maintenance service", "data processing service" and other sub-classes, clearly defining the attributes and relationships of each sub-class, and building the basic framework of the ontology;

[0019] S112. From bottom to top, using natural language processing (NLP) technology and data mining algorithm, the collected multi-source data is analyzed, new technical concepts are extracted from technical documents, and service demand related concepts are mined from user reviews. These newly discovered concepts and relationships are gradually supplemented into the pre-constructed ontology framework, making the ontology more in line with actual business needs.

[0020] Further, in the S1, the specific steps of collecting multi-source data and encoding into an initial graph by a graph neural network include:

[0021] S121. Definition of nodes and edges: define entities such as services, users, and technologies as nodes, and define edges according to the business relationship between entities. If user A uses service B, a "use" relationship edge is established between user A node and service B node.

[0022] S122. Information transmission and aggregation, GNN transmits information through multiple layers, allowing nodes and neighbor nodes to exchange information. In each layer, nodes aggregate neighbor node information according to edge weights, and update their own representation combining their own characteristics. Finally, multi-source data is converted into an initial graph containing rich semantic information.

[0023] Further, in the S1, the method for establishing static and dynamic evolution characteristics of service capabilities is established.

[0024] S131. Application of static ZIP model, extract static characteristics from basic attributes and fixed resource allocation of services, and quantitatively represent them through ZIP model to form a static description of service capabilities.

[0025] S132. Application of dynamic induction motor model, real-time monitoring of dynamic data such as service performance indicators and resource dynamic allocation; analog operation characteristics of dynamic induction motor, establish mathematical model to model the dynamic evolution of service capabilities, and capture the change rule of service performance over time.

[0026] Further, in the S1, the step of using multi-head attention mechanism to filter high correlation entities to form a service network includes:

[0027] S141. Multi-head parallel computing, multiple "heads" are used for parallel computing, each "head" captures the relationship between entities from different angles; for example, one "head" focuses on the association between services and technologies, and another "head" focuses on the matching degree between services and user needs.

[0028] S142. Key information filtering, integrate the calculation results of multiple "heads", assign attention weights to each entity, filter out entities highly related to service capabilities, remove redundant information, and construct a service network focusing on core business relationships, providing a refined data foundation for subsequent operations.

[0029] Further, in S2, the method of multi-modal demand feature dynamic extraction is as follows:

[0030] S211. Multi-modal data access and preprocessing:

[0031] User demand may be presented in various modalities such as text, image, voice, etc. First, collect and preprocess multi-modal data:

[0032] Text data, perform word segmentation, part-of-speech tagging, remove stop words, etc. Operations convert natural language text into computer processable word vector form;

[0033] Image data, extract key features such as color, shape, object category, etc. from image recognition technology, and convert to feature vector;

[0034] Speech data, convert speech content to text, and perform the same preprocessing operations as text data;

[0035] S212. Improved SHAP model analysis, use the improved SHAP model to analyze the features of the preprocessed multi-modal data;

[0036] Feature contribution calculation, based on game theory, calculate the contribution value of each feature to the model output; When analyzing user demand for software interface, determine the contribution degree of interface color, button layout, etc. to user satisfaction prediction through SHAP model, and identify key demand factors;

[0037] Key feature screening, according to the size of the feature contribution value, set a threshold to screen out key features that have a greater impact on user demand, reduce data dimension, and improve subsequent processing efficiency;

[0038] S213. Cross-modal attention mechanism application, use cross-modal attention mechanism to mine potential links between multi-modal data;

[0039] Inter-modal correlation establishment, calculate attention weights between feature vectors of different modalities such as text, image, and voice, and determine the importance of each modality information;

[0040] Structured feature construction, fuse multi-modal features processed by attention mechanism, construct a unified structured feature representation, and build an evolution graph based on it to intuitively show the trend of user demand over time;

[0041] S214. Demand change node identification, in the evolution graph, determine the demand change by setting a suitable threshold;

[0042] Threshold setting, determine the threshold of node change degree according to historical data and business experience;

[0043] Change node determination, periodically calculate the change degree of each node, compare with the threshold value; when the node change exceeds the threshold value, mark it as a demand change node, and generate a demand vector containing key demand characteristics according to these nodes, which provides the basis for service matching.

[0044] Further, the S3 comprises:

[0045] S311. Similarity calculation, similarity calculation between demand vector and service feature vector in initial graph, similarity calculation is:

[0046] ;

[0047] Wherein, is the similarity degree of demand and service; is the demand vector; is the service feature vector; by calculating the cosine value of the angle between the two vectors, the similarity degree of demand and service is measured, the value is closer to 1, indicating that the matching degree is higher;

[0048] S312. Introducing double DQN network to realize dynamic matching strategy adjustment, according to the similarity calculation result and environmental feedback, the matching strategy is adjusted adaptively; two neural networks with the same structure but different parameter updates are constructed, one is used to evaluate the action value, and the other is used to select the action. In the service matching scene, "action" is to select different service combinations for matching; by constantly interacting with the environment, the network parameters are updated according to the reward mechanism, so that the double DQN network can adaptively adjust the matching strategy according to the demand change, and avoid overestimation;

[0049] S313. Adam optimizer generates objective function, Adam optimizer adjusts learning rate adaptively according to gradient history information of different parameters; in the matching process, give larger learning rate to frequently changing parameters to speed up convergence; give smaller learning rate to relatively stable parameters to ensure the accuracy of optimization; consider multiple factors such as matching accuracy, service cost, user satisfaction, etc. to build objective function; the objective function is represented as:

[0050] ;

[0051] is the objective function value of Adam optimizer; is the weight coefficient in the objective function of Adam optimizer, used to adjust the importance of each factor in the objective function, Adam optimizer adjusts the learning rate adaptively according to the gradient history information of each parameter, and finds the parameter combination that makes the objective function value maximum; is the service use cost; is the user satisfaction; The matching accuracy refers to the degree to which the matching results conform to the user's actual needs;

[0052] S314. Adversarial sample verification: Design disturbances are made to the original demand data and service data, including adding interference words to the demand text, modifying service feature parameters, etc., to generate adversarial samples; the adversarial samples are input into the matching scheme and the changes in the matching results are observed. If the matching scheme can still maintain a high accuracy and stability under the adversarial sample, it means that it has strong robustness. The verified scheme forms a multi-target matching scheme.

[0053] Furthermore, in S4, the step of estimating the difference using a kernel function includes:

[0054] Based on the service and demand matching results obtained in step S3 and the feedback data from users after actually using the service, the kernel function is used to perform difference estimation, and the radial basis function (RBF) is used to calculate the difference between the predicted value and the actual value. The formula is:

[0055] ;

[0056] is the difference between the predicted value and the actual value; is the prediction vector; is the actual feedback vector; It is set according to the fluctuation of historical data; The closer the value is to 0, the greater the difference between the predicted result and the actual situation. When it is less than the set threshold, the graph iteration process is triggered to optimize the service characteristics and matching algorithm.

[0057] Furthermore, in S4, the dynamic load modeling step includes:

[0058] S411. Data Collection and Problem Diagnosis: After confirming significant discrepancies, we initiated in-depth collection of service operation data, including real-time performance metrics, service call logs, and resource allocation records. Analysis of data from high-concurrency periods revealed that CPU utilization remained consistently above 90%, memory usage remained above 85%, and response time and throughput fluctuated significantly. Combined with frequent resource request timeouts in service logs, we determined that a resource bottleneck existed.

[0059] S412. Modify the static characteristics of the service based on service performance issues and user feedback. If the service does not meet the promised service level, downgrade the service level attributes. If the service is found to have defects in specific scenarios, add new tags to the fixed resource configuration attributes and provide detailed descriptions to improve the basic attribute description of the service.

[0060] S413. Use exponential smoothing method to predict service performance indicators. The formula is as follows:

[0061] ;

[0062] wherein, current cycle prediction response time; last cycle prediction value; is the actual measurement value of the current cycle; is the smoothing coefficient, which reflects the degree of dependence on historical data; The closer to 0, the more the model relies on historical prediction values; The closer to 1, the more emphasis on the current measured value;

[0063] The updated response time prediction value and the changes in resource usage such as CPU, memory, etc. are synchronized to the dynamic feature vector of the service in the knowledge graph, ensuring that the service features are real-time synchronized with the actual running state.

[0064] Further, in the S4, the specific steps of improving the matching ability of the closed loop mechanism are:

[0065] S421. Entity relationship adjustment, based on service feature update and historical optimization experience, analyze the association between service and technology nodes. If a certain technology has a significant effect on service performance, add a relationship edge between the corresponding nodes, and give a weight value between 0 and 1 according to the historical optimization effect evaluation. The higher the weight, the closer the association between the two;

[0066] At the same time, according to the satisfaction feedback of users after using the service, dynamically adjust the weight of the "use" relationship between the user node and the service node. When the satisfaction decreases, the weight value is adjusted downward in proportion; When the satisfaction improves, the weight value is adjusted upward accordingly, so as to more accurately depict the actual relationship between the service and the user;

[0067] S422. Ontology extension, extract new demand concepts from user feedback text through natural language processing technology, and supplement them to the ontology structure of the knowledge graph by using the bottom-up method; According to the semantics and business logic of the new demand concept, associate it to the appropriate ontology subclass;

[0068] At the same time, define the related attributes of the new demand concept, and clearly define its features and constraints, so as to extend and enrich the ontology structure of the knowledge graph, so that it can better cover the diversified needs of users, and provide comprehensive semantic support for matching;

[0069] S423. Matching algorithm optimization, double DQN network parameter adjustment, in order to make the double DQN network more suitable for actual demand, adjust the reward function; On the basis of the original reward function, add a penalty term related to the performance of the service peak period, and the new reward function is:

[0070] ;

[0071] for balancing the similarity parameter; for the cost parameter; for the satisfaction parameter; for the peak performance parameter; is positively correlated with the degree of exceeding the service peak response time. When the response time of the service in the peak period exceeds the preset threshold value, the value of increases, resulting in the overall reward value decreases;

[0072] S424. Adversarial sample enhancement. To ensure the robustness of the matching algorithm in extreme scenarios, adversarial samples simulating extreme loads are generated, which are input into the optimized matching model to check whether the matching scheme output by the model is effective. By setting evaluation indicators such as matching accuracy, the matching algorithm parameters are continuously adjusted and optimized to ensure that the model can provide reliable matching results when facing various complex and extreme user demands and service scenarios, improving the stability and reliability of the system;

[0073] S425. New matching application. When a user submits a service demand, the optimized knowledge graph and matching algorithm work together; the knowledge graph filters out a service set that matches the user's demand in terms of semantics and technical association based on the updated entity relationship and ontology structure; the matching algorithm further selects the most suitable service for the user's demand from the filtered service set by combining the dynamic characteristics of the service, such as response time, throughput, and other performance indicators, to generate a matching scheme;

[0074] S426. Closed-loop feedback mechanism. The new matching result and user feedback will enter the S4 link again, repeating the difference estimation, service feature update, and algorithm optimization process. Through continuous cyclic iteration, the knowledge graph continuously enriches and corrects the service feature information, and the matching algorithm continuously optimizes the strategy, ultimately realizing the spiral improvement of the accuracy and adaptability of intelligent matching of information technology services.

[0075] In summary, due to the adoption of the above technical solutions, the beneficial technical effects of the invention are:

[0076] The deep integration of low carbonization and resource utilization is achieved by embedding low carbon goals into service matching processes, combining static resource configuration modeling with dynamic performance tracking, and realizing the quantitative management of service running energy efficiency. Specifically, the static ZIP model is used to depict the service basic attributes and fixed resources, and the dynamic induction motor model is used to capture dynamic data such as CPU usage and response time in real time, so that the system can preferentially select low-energy and high-resource efficient service instances. This mechanism breaks the traditional "heavy function, light energy efficiency" mode, effectively avoids the invalid call of high-energy consumption services, promotes the deep integration of information technology services and green computing trends, and provides technical support for the low carbon operation of power distribution networks and other scenarios.

[0077] Comprehensive understanding and dynamic response of multi-modal demand, in view of the limitations of traditional single-modal demand analysis, the invention realizes the integrated analysis of multi-form demands such as text, image and voice by improving SHAP model and cross-modal attention mechanism. For example, the text description, interface design sketch and voice instruction in user demand are converted into unified structured feature vectors, and the evolution graph is used to track the demand change nodes in real time. This ability enables the system to accurately capture the user's potential demand and dynamically adjust the matching strategy based on the double DQN network, significantly improving the timeliness and accuracy of demand response, and avoiding the decline of service quality caused by one-sided or lagging demand understanding.

[0078] Systematic breakthrough of algorithm efficiency and robustness, the construction and optimization mechanism of knowledge graph provides a semantic-rich resource base for intelligent matching, effectively solving the problem of single feature representation of traditional algorithms. Through cosine similarity calculation, multi-objective function construction of Adam optimizer and adversarial sample verification, the system can realize adaptive adjustment of matching strategy and improvement of anti-interference ability in the scene of rapid change of multi-modal demand and spatio-temporal evolution of service features. Specifically, it can quickly filter out high-correlation candidate sets from a large number of services, balance the matching accuracy, service cost and user satisfaction, and maintain the stability of the matching results in the presence of abnormal data or extreme load scenarios.

[0079] Continuous evolution ability driven by closed-loop iteration, based on the difference estimation of matching results and user feedback, the system can trigger dynamic load modeling and knowledge graph iteration process. By updating the service feature vector, adjusting the entity relationship weight and expanding the ontology structure, the system realizes the continuous optimization of service capability representation and matching algorithm. This closed-loop mechanism of "matching-feedback-evolution" enables the system to continuously adapt to the development of industry technology and changes in user demand, avoiding the inefficiency of traditional methods relying on manual rule modification, and forming a self-improving and spiral-enhancing intelligent ecosystem.

[0080] The technical scheme of the application is not the independent application of a single technology, but the synergistic effect of multiple modules such as knowledge graph, machine learning, and optimization algorithm, forming a comprehensive response capability to the dynamic complexity of information technology services. For example, in the power distribution network scenario, multiple goals such as low-carbon scheduling, multi-modal demand analysis, and dynamic resource allocation during peak periods can be met simultaneously, and the matching efficiency is continuously improved through closed-loop iteration. This systematic innovation makes the application not only suitable for traditional information technology service scenarios, but also provides efficient and flexible intelligent matching solutions for emerging fields such as energy internet and smart grid. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 It is a logic block diagram of an information technology service intelligent matching method based on knowledge graph driving. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical scheme and advantages of the invention more clear and explicit, the invention will be further described in detail below combined with embodiments. It should be understood that the specific embodiments described herein are only used to explain the invention and do not limit the invention.

[0083] As shown in Figure 1 A knowledge graph driven information technology service intelligent matching method, comprising the following steps:

[0084] S1. Multi-source heterogeneous data knowledge graph construction, combining top-down and bottom-up methods to construct the ontology of information technology service field, collecting multi-source data and encoding into initial graph by graph neural network;

[0085] Based on the parameter identification method of static ZIP model and dynamic induction motor model, the static characteristics and dynamic evolution characteristics of service capability are established, the high correlation entities are selected by multi-head attention mechanism, the service network is formed, and the resource foundation is provided for the subsequent steps;

[0086] S2. Multi-modal demand feature dynamic extraction, based on the initial graph formed in S1, the multi-modal feature of user demand is extracted by improved SHAP model and cross-modal attention mechanism, the evolution graph is constructed, the demand change node is identified by threshold comparison, and the demand vector is generated;

[0087] S3. Service and demand dynamic matching optimization, calculating the similarity of demand vector in S2 and initial graph feature in S1, adjusting the matching strategy by double DQN network, generating the objective function by Adam optimizer, verifying by adversarial samples, forming a multi-objective matching scheme;

[0088] S4. Matching result driven atlas iteration, based on S3 matching results and user feedback, use kernel function to estimate difference, update service features according to dynamic load modeling, improve matching ability through optimization closed loop mechanism, the optimized atlas is used for new matching, forming a cycle.

[0089] In S1, the specific steps of constructing the ontology of the information technology service field combining top-down and bottom-up methods are as follows:

[0090] S111. Top-down design, organize field experts and technical personnel, according to information technology service industry standards and business logic, sort out the basic concept framework, define "information technology service" as the top class, with "software development service", "system operation and maintenance service", "data processing service" and other sub-classes, and clearly define the attributes (such as service period, service price) and relationships (such as the support relationship between service and technology) of each sub-class, and build the basic framework of the ontology;

[0091] S112. Bottom-up improvement, using natural language processing (NLP) technology and data mining algorithms, in-depth analysis of the collected multi-source data, in technical documents, using named entity recognition technology, extracting new technical concepts (generative AI development framework, edge computing solution, etc.) from the text; in user reviews, through sentiment analysis and semantic understanding, mining concepts related to service demand, such as users mentioning "hope the service has stronger real-time data processing capability", thus extracting the demand concept of "real-time data processing". Gradually supplement these newly discovered concepts and relationships to the pre-constructed ontology framework. If a new technology is found to be adopted by multiple services, establish the relationship between the technology and the related service sub-classes in the ontology, so that the ontology is constantly enriched and improved, and more closely meets the actual business needs.

[0092] In S1, the specific steps of encoding the collected multi-source data into the initial atlas by graph neural network include:

[0093] After the ontology is constructed, multi-source data cannot be directly used for knowledge graph construction due to its diversity in source and format, and needs to be preprocessed, mainly including data cleaning and format conversion. In the data cleaning link, duplicate data is removed by duplicate checking algorithm, and incorrect data is corrected by data checking rules; for unstructured data, text parsing, image recognition, speech to text and other technologies are used to convert it into structured data. Encode using graph neural network (GNN), the specific process is as follows:

[0094] S121. Definition of nodes and edges, various entities in the field of information technology services are defined as nodes, such as "cloud computing services", "a certain enterprise user", "Python programming language", etc. as service nodes, user nodes and technology nodes. According to the actual business contact between entities, edges are defined. If user A uses service B, an edge of "use" relationship is established between user A node and service B node; if the implementation of service B depends on technology C, an edge of "dependence" relationship is established between service B node and technology C node. In this way, entities and relationships are converted into basic elements of graph structure.

[0095] S122. Information transmission and aggregation, GNN realizes information exchange between nodes and neighbor nodes through multi-layer propagation mechanism; in each layer, nodes will aggregate the information of neighbor nodes according to the weight of the edge connected with them; the weight of the edge reflects the closeness of the relationship between nodes; if a user frequently uses a service, the weight of the edge between the user node and the service node will be higher; after aggregating the information of neighbor nodes, the node updates its own representation through a specific calculation method combined with its own existing characteristics; after multi-layer information transmission and aggregation, the multi-source data is finally converted into an initial graph containing rich semantic information, so that each node in the graph contains its own characteristics and comprehensive information after interacting with neighbor nodes.

[0096] In S1, the representation method of static characteristics and dynamic evolution characteristics of service capability is established:

[0097] In order to accurately describe the service capability, static ZIP model and dynamic induction motor model are introduced to model the service capability from static and dynamic dimensions.

[0098] S131. Application of static ZIP model, static characteristics are extracted from the basic attributes and fixed resource configuration of services. Taking cloud storage service as an example, its basic attributes include service type (such as object storage, block storage), service level (ordinary level, enterprise level); fixed resource configuration involves server quantity, storage capacity, supported file types, etc. Through ZIP model, these static characteristics are quantitatively represented, such as converting storage capacity, supported file types and other information into numerical vectors to form a static description of service capability. This static description provides a basic attribute basis for subsequent service matching, which can quickly filter out services that meet the basic conditions;

[0099] S132. Dynamic induction motor model application: real-time monitoring of performance indicators and resource dynamic allocation during service operation, such as service response time, throughput, CPU usage change, memory allocation adjustment, etc. Analogous to the operating characteristics of dynamic induction motors, a mathematical model is established to model the dynamic evolution of service capacity. Dynamic induction motors adjust their speed, torque and other parameters dynamically as the load changes, and service capacity is also affected by factors such as user request volume and business complexity. By establishing a mathematical model, the changing pattern of service performance over time is captured, such as predicting the response time trend of the service in the future, thereby providing decision support for dynamic allocation of service resources and optimization of service quality.

[0100] In S1, the step of screening high correlation entities using multi-head attention mechanism to form a service network includes:

[0101] S141 Multi-head parallel computing, multi-head attention mechanism calculates in parallel through multiple "heads", each "head" captures the relationship between entities from different angles. For example, one "head" focuses on analyzing the correlation between services and technologies, and mining the core technologies on which services depend and the impact of technologies on service performance; another "head" focuses on the matching degree of service and user demand, and analyzes the fit between user demand characteristics and service providing capacity; each "head" independently calculates the attention weight between entities, and extracts relationship information between entities from different dimensions;

[0102] S142. Key information screening, the calculation results of multiple "heads" are integrated to assign a comprehensive attention weight to each entity in the initial graph. According to the weight size, entities highly related to service capacity are screened out, and redundant entities and relationships with small impact on service capacity are removed. For example, among numerous technology nodes, core technology nodes that play a key role in service performance and service nodes that match high-frequency user demand are selected. In this way, a service network focusing on core business relationships is constructed, which retains the most important entity and relationship information, provides a refined data foundation for subsequent demand feature extraction and service matching optimization, and can effectively improve the calculation efficiency and matching accuracy.

[0103] In S2, the method for dynamically extracting multi-modal demand features includes:

[0104] S211. Multi-modal data access and preprocessing:

[0105] User demand presents a multi-modal characteristic, which may be presented in the form of text, image, voice, etc. First, these multi-modal data need to be collected and preprocessed for subsequent analysis;

[0106] Text data, for user-submitted requirement description documents and other text data, is processed using natural language processing techniques; first, a word segmentation operation is performed to divide continuous text into individual words; then, part-of-speech tagging is performed to identify the part of speech (noun, verb, adjective, etc.) of each word; finally, stop words such as "of", "was", "in" and other words that have little impact on semantic understanding are removed; through the Word2Vec model, natural language text is converted into a computer-processable word vector form, allowing text information to be calculated and analyzed in the form of numerical vectors;

[0107] Image data, for user-provided interface design diagrams and other image data, uses image recognition techniques to extract key features; the ResNe model of the convolutional neural network is used to extract features from the image, automatically learning features such as color, shape, and object category, and converting them into feature vectors; for example, the shape, color distribution, and interface layout of the button are extracted from the interface design diagram, and the feature information is represented in the form of a numerical vector, providing data support for subsequent analysis of user interface design requirements;

[0108] Voice data, using a Transformer-based speech recognition model to convert speech content into text, after converting speech into text, the same preprocessing operations as text data are performed, such as word segmentation, part-of-speech tagging, removing stop words, and word embedding, to convert speech information into a computer-processable vector form;

[0109] S212. Improved SHAP model analysis, using an improved SHAP (SHapley Additive exPlanations) model to analyze the features of the preprocessed multi-modal data in depth to identify key demand factors;

[0110] Feature contribution calculation, based on game theory principles, the improved SHAP model calculates the contribution value of each feature to the model output (user demand judgment), the improved SHAP model considers each feature as a participant in the game, and the SHAP model quantifies the contribution of each feature to the final model output result by simulating all possible feature combinations; the specific calculation process is as follows: for the input feature set , represents the feature, is the total number of features; these features include color parameters, button layout coordinates, and function module types of the software interface after multi-modal information fusion, and the model prediction result is , is a user satisfaction prediction model based on deep learning; the contribution value of feature is calculated as follows:

[0111] ;

[0112] Representation characteristics Contribution value; is a feature subset; is a feature subset size; It is a weight coefficient used to balance the influence of subsets of different sizes when calculating the contribution value; It means adding features The change in the output of the model before and after, the larger the change, the better the feature The more significant the impact on the model output;

[0113] Key feature screening: After obtaining the contribution value of each feature, set a threshold based on the size of the feature contribution value to screen out key features that have a greater impact on user needs. By screening key features, the data dimension is reduced, redundant features with less impact on user needs are removed, and subsequent processing efficiency is improved, allowing the model to focus more on analyzing and matching key demand factors;

[0114] S213. Application of cross-modal attention mechanism: using cross-modal attention mechanism to explore potential connections between multimodal data and achieve fusion and structured representation of multimodal information;

[0115] Inter-modal associations are established, and attention weights are calculated between feature vectors of different modalities such as text, images, and voice to determine the importance of each modal information. When the user provides both a text description of requirements and an interface design sketch, the cross-modal attention mechanism calculates the attention weights between the text feature vector and the image feature vector. For example, if the text describes "the search button needs to be highlighted", the interface elements related to the search button are found in the image feature vector. By calculating the attention weights, the degree of association between the text description and the corresponding elements in the image is determined, thereby finding the correspondence between the text description and the image elements and achieving alignment of information from different modalities.

[0116] Among them, the formula for the text-image modality attention weight is:

[0117] ;

[0118] in, and It is a learnable weight matrix used to map text and image feature vectors to a suitable space to better calculate similarity. Its dimension is determined by the actual task and feature vector dimension. For example, if the dimension is 768, in order to map it to a dimension suitable for attention calculation, the dimension may be set to 768×64; is the scaling factor, is the dimension of the mapping vector, this factor is introduced to avoid the result of inner product being too large function gradient disappears, ensuring the stability of weight calculation; The function normalizes the calculation result and converts it into a probability distribution form, so that the sum of attention scores is 1, and the is the attention weight between the text and image feature vectors, reflecting the relative importance of text and image information in the current demand understanding; is the text feature vector, obtained by preprocessing and feature extraction of the text description of user demand; is the image feature vector, obtained by image recognition technology processing of the image provided by the user;

[0119] In practical application, when the user provides both the text demand description "design a simple blue main interface containing a centered search button" and the interface design sketch, the system converts the text description into a text feature vector through natural language processing technology, and the sketch is processed into an image feature vector through image recognition technology. Then the attention weight between the two is calculated using the above formula; if the key descriptions such as "blue main interface" and "search button" have corresponding elements in the image, and the calculated attention weight is high, it means that these text descriptions are closely related to the corresponding elements in the image and have an important position in demand understanding.

[0120] Similarly, the attention weight between text-speech, image-speech and other modal combinations can be calculated to comprehensively measure the importance of information in each modality.

[0121] Structural feature construction, using weighted summation to fuse multi-modal features processed by attention mechanism, key features of text, image, speech and other modalities are fused according to their importance to form a structured comprehensive vector containing multi-modal information. The evolutionary graph is constructed based on the structured feature representation, the nodes in the graph represent different demand features, such as the color of the software interface, the button layout, the function module, etc. Each node corresponds to a part of the information in the structured feature vector; the edges represent the relationship between the features, and the weight of the edge is determined by the attention score between the features; for example, if the attention score between the "search button" described in the text and the corresponding button element in the image is high, then the weight of the edge between the node representing the "search button" text description and the node representing the button element in the image in the evolutionary graph is large, indicating that these two features are closely related.

[0122] S214. Demand change node identification, in the evolutionary graph, set a suitable threshold to judge the demand change and capture the dynamic changes of user demand in time;

[0123] Threshold setting: according to historical data and business experience, the threshold of the degree of node change is determined; for example, through analysis of a large amount of historical demand data, it is found that when the change amplitude of a certain demand feature exceeds 15%, it usually means that the user demand has changed significantly, so 15% is set as the threshold; the setting of the threshold needs to consider the industry characteristics, business scenarios and data fluctuation to ensure that the real demand change can be accurately identified;

[0124] Change node determination: periodically calculate the change degree of each node in the evolution graph, which is realized by comparing the difference of feature vectors of nodes at different time points, and the calculation formula of the change degree of each node in the evolution graph is:

[0125] ;

[0126] Among them, represents the change degree of each node in the evolution graph; is the dimension of the feature vector; represents the first component in the feature vector; and represent the first component value in the feature vector at time , ; When is a preset threshold, mark the node as a demand change node;

[0127] According to these demand change nodes, extract the key demand features contained therein to generate demand vectors, which provide dynamic and accurate basis for subsequent service and demand matching, and ensure that services can respond to changes in user demand in a timely manner.

[0128] The step of service and demand dynamic matching optimization in S3 includes:

[0129] S311. Similarity calculation: calculate the similarity between the demand vector and the service feature vector in the initial graph, and the commonly used cosine similarity formula is:

[0130] ;

[0131] Among them, is the similarity degree of demand and service; is the demand vector; is the service feature vector; by calculating the cosine value of the angle between the two vectors, the similarity degree of demand and service is measured, and the value is closer to 1, indicating higher matching degree; through similarity calculation, the candidate services that are more matched with user demand can be screened from the numerous services in the initial graph;

[0132] S312. Introduce a double DQN network to realize dynamic matching strategy adjustment. According to the similarity calculation result and environmental feedback, the matching strategy is adjusted adaptively; the double DQN network includes an evaluation network and a target network, which learns the optimal strategy by constantly interacting with the environment (i.e., trying different matching schemes and obtaining feedback);

[0133] During the process, two neural networks with the same structure but different parameter updates are constructed, one for evaluating action value and one for selecting action; in the service matching scenario, "action" means selecting different service combinations for matching; the double DQN network updates the network parameters according to the reward mechanism by constantly interacting with the environment, so that the double DQN network can adaptively adjust the matching strategy according to the demand change and avoid overestimation problem;

[0134] The reward function considers factors such as matching accuracy, service cost, and user satisfaction, and the formula is:

[0135] ;

[0136] is the reward function value; is the balance similarity parameter; is the cost parameter; is the satisfaction parameter; the weight coefficient in the double DQN network reward function; is the service usage cost; is the user satisfaction;

[0137] The double DQN network constantly tries different matching actions and updates the network parameters according to the reward feedback. The evaluation network predicts the expected reward value of each action, and the target network is used to stabilize the learning process and periodically copies the parameters from the evaluation network. After multiple rounds of iterative learning, the double DQN network adjusts the matching strategy and preferentially selects services with higher comprehensive rewards for matching;

[0138] S313. The Adam optimizer generates the objective function. The Adam optimizer adaptively adjusts the learning rate according to the gradient history information of different parameters; in the matching process, a higher learning rate is given to frequently changing parameters to speed up convergence; a smaller learning rate is given to relatively stable parameters to ensure the accuracy of optimization; multiple factors such as matching accuracy, service cost, and user satisfaction are considered to construct the objective function; the objective function is represented as:

[0139] ;

[0140] is the objective function value generated by the Adam optimizer; The weight coefficient in the Adam optimizer's objective function is used to adjust the importance of each factor in the objective function. The Adam optimizer adaptively adjusts the learning rate based on the historical gradient information of the objective function for each parameter, and seeks the parameter combination that maximizes the objective function value. The matching accuracy refers to the degree to which the matching results meet the user's actual needs. During the optimization process, the Adam optimizer continuously adjusts the parameters in the matching strategy (such as service selection priority, weight distribution, etc.) so that the generated matching solution achieves the optimal balance between multiple objectives;

[0141] S314. Adversarial sample verification: Design disturbances are made to the original demand data and service data, including adding interference words to the demand text, modifying service feature parameters, etc., to generate adversarial samples; the adversarial samples are input into the matching scheme and the changes in the matching results are observed. If the matching scheme can still maintain a high accuracy and stability under the adversarial sample, it means that it has strong robustness. The verified scheme forms a multi-target matching scheme.

[0142] In S4, the step of estimating the difference using the kernel function includes:

[0143] Based on the service and demand matching results obtained in step S3 and the feedback data after users actually use the service, the kernel function is used to perform difference estimation, and the radial basis function (RBF) is used to calculate the difference between the predicted value and the actual value. The formula is:

[0144] ;

[0145] is the difference between the predicted value and the actual value; is the prediction vector; is the actual feedback vector; It is set according to the fluctuation of historical data; The closer the value is to 0, the greater the difference between the predicted result and the actual situation. When it is less than the set threshold, the graph iteration process is triggered to optimize the service characteristics and matching algorithm.

[0146] In S4, the dynamic load modeling step includes:

[0147] S411. Data collection and problem diagnosis. After confirming significant discrepancies, we initiated in-depth collection of service operation data, covering real-time performance metrics (CPU usage, memory utilization, disk I / O, network bandwidth), service call logs, and resource allocation records. Analysis of data during high-concurrency periods (such as peak business hours) revealed that CPU utilization remained consistently above 90%, memory utilization remained above 85%, and response time and throughput showed significant fluctuations. Combined with frequent resource request timeout records in service logs, we determined that a resource bottleneck existed.

[0148] S412. According to the service performance problems and user feedback, the static characteristics of the service are corrected; if the service does not reach the promised service level, the service level attribute is lowered; if it is found that the service has defects in specific scenarios, a label (such as "resource shortage in high concurrency scenario") is added in the fixed resource configuration attribute, and detailed explanations are supplemented to perfect the basic attribute description of the service;

[0149] S413. Exponential smoothing method is used to predict the service performance index, and the formula is as follows:

[0150] ;

[0151] Wherein, the predicted response time in the current period; the predicted value in the last period; is the actual measured value in the current period; the smoothing coefficient (value range 0-1), which reflects the degree of dependence on historical data; The closer to 0, the more the model depends on the historical predicted value; The closer to 1, the more emphasis on the current measured value;

[0152] The updated response time prediction value and the changes in resource usage such as CPU and memory are synchronized to the dynamic feature vector of the service in the knowledge graph, ensuring that the service features are real-time synchronized with the actual running state.

[0153] In the S4, the specific steps of the closed-loop mechanism to improve the matching ability are as follows:

[0154] S421. Entity relationship adjustment, based on service feature update and historical optimization experience, analyze the association between service and technology nodes. If a certain technology has a significant improvement effect on service performance, add a relationship edge between the corresponding nodes, and according to the historical optimization effect evaluation, give a weight value between 0 and 1. The higher the weight, the closer the association between the two;

[0155] At the same time, according to the satisfaction feedback of users after using the service, dynamically adjust the weight of the "use" relationship between the user node and the service node. When the satisfaction decreases, the weight value is lowered in proportion; when the satisfaction improves, the weight value is increased accordingly, so as to more accurately depict the actual relationship between the service and the user;

[0156] S422. Ontology extension, extract new demand concepts from user feedback text through natural language processing technology, and supplement them to the ontology structure of the knowledge graph by using the bottom-up method. According to the semantics and business logic of the new demand concept, it is associated to the appropriate ontology subclass;

[0157] Meanwhile, the related attributes of the new demand concept are defined to clarify its characteristics and constraints, thus extending and enriching the ontology structure of the knowledge graph, making it better cover the diversified demands of users and provide comprehensive semantic support for matching;

[0158] S423. Matching algorithm optimization, double DQN network parameter adjustment. To make the double DQN network more suitable for actual demand, the reward function is adjusted. On the basis of the original reward function, a penalty term related to the performance of service peak period is added. The new reward function is:

[0159] ;

[0160] the performance parameter of the peak period; is positively correlated with the degree of exceeding the response time of the service peak period. When the response time of the service in the peak period exceeds the preset threshold, increases, resulting in a decrease in the overall reward value ;

[0161] S424. Adversarial sample enhancement. To ensure the robustness of the matching algorithm in extreme scenarios, adversarial samples simulating extreme loads are generated, such as adjusting key indicators such as request volume and data traffic to far exceed normal levels. These adversarial samples are input into the optimized matching model to check whether the matching scheme output by the model is effective. By setting evaluation indicators such as matching accuracy, the matching algorithm parameters are continuously adjusted and optimized to ensure that the model can provide reliable matching results when facing various complex and extreme user demands and service scenarios, improving the stability and reliability of the system;

[0162] S425. New matching application. When users submit service demands, the optimized knowledge graph and matching algorithm work together; the knowledge graph filters out a set of services that are semantically and technically related to user demands based on updated entity relationships and ontology structure. The matching algorithm further selects the most suitable service for user demands from the filtered service set by combining dynamic characteristics of services such as response time and throughput, and generates a matching scheme;

[0163] S426. Closed-loop feedback mechanism. The new matching results and user feedback will again enter the S4 link, repeating the difference estimation, service feature update, and algorithm optimization process. Through this continuous cycle iteration, the knowledge graph continuously enriches and corrects service feature information, and the matching algorithm continuously optimizes strategies, ultimately achieving spiral improvement in the accuracy and adaptability of intelligent matching of information technology services.

[0164] The above describes the preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A knowledge graph-driven intelligent matching method for information technology services, characterized in that: The following steps are involved: S1. Constructing a knowledge graph for multi-source heterogeneous data, combining top-down and bottom-up approaches to build an ontology for the information technology service field. Collecting multi-source data and encoding it into an initial graph using a graph neural network. Based on the parameter identification method of the static ZIP model and the dynamic induction motor model, the static and dynamic evolution characteristics of service capabilities are established. The multi-head attention mechanism is used to screen highly relevant entities to form a service network, providing a resource foundation for subsequent steps. S2. Dynamic extraction of multimodal demand features. Based on the initial graph formed in S1, an improved SHAP model and a cross-modal attention mechanism are used to extract multimodal features of user demands. An evolution graph is constructed, and demand change nodes are identified through threshold comparison. A demand vector is generated as the key basis for matching. S3. Dynamically optimize the matching of services and demands. This involves calculating the similarity between the demand vector in S2 and the initial graph features in step 1. The matching strategy is adaptively adjusted using a dual DQN network. The Adam optimizer generates the objective function, which is then validated using adversarial examples to form a multi-objective matching solution. S4. Matching results drive graph iteration. Based on S3 matching results and user feedback, kernel functions are used to estimate differences. Service characteristics are updated based on dynamic load modeling. The matching algorithm is improved through a closed-loop optimization mechanism. The optimized graph is then used for new matching, forming a cycle. In S2, the method for dynamically extracting multimodal demand features is as follows: S211. Multimodal data access and preprocessing: User needs are presented in multiple modalities such as text, images, and voice. Multimodal data is collected and preprocessed: S212. Improved SHAP model analysis: Use the improved SHAP model to analyze the preprocessed multimodal data features; Feature contribution calculation, based on game theory principles, calculates the contribution of each feature to the model output. When analyzing user requirements for the software interface, the SHAP model is used to determine the contribution of interface color and button layout features to user satisfaction prediction, identifying key demand factors. Key feature screening: according to the size of the feature contribution value, set the threshold to screen out the key features that have a greater impact on user needs, reduce data dimensions, and improve subsequent processing efficiency; S213. Application of cross-modal attention mechanism, using cross-modal attention mechanism to explore potential connections between multimodal data; Establish inter-modal associations, calculate attention weights between feature vectors of different modalities such as text, image, and speech, and determine the importance of each modal information; Structured feature construction: integrating multimodal features processed by the attention mechanism to build a unified structured feature representation. Based on this, an evolution graph is constructed to intuitively show the changing trends of user needs over time. S214. Identify demand change nodes. In the evolution graph, determine demand changes by setting appropriate thresholds. Threshold setting: Determine the threshold of node change degree based on historical data and business experience; The change nodes are determined, and the degree of change of each node is calculated regularly and compared with the threshold. When the node change exceeds the threshold, it is marked as a demand change node, and a demand vector containing key demand features is generated based on the node to provide a basis for service matching.

2. The method for intelligent matching of information technology services based on knowledge graph drive according to claim 1 is characterized in that: In S1, the specific steps of collecting multi-source data and encoding them into an initial graph through a graph neural network include: S121. Definition of nodes and edges: Define services, users, and technical entities as nodes, and define edges based on the business connections between entities; S122. Information transmission and aggregation. GNNs enable nodes to exchange information with neighboring nodes through multi-layer propagation. In each layer, nodes aggregate information from neighboring nodes based on the weights of their connected edges and update their own representations based on their own characteristics. Ultimately, multi-source data is transformed into an initial graph containing rich semantic information. In S1, the static characteristics and dynamic evolution characteristics of service capabilities are represented as follows: S131. Static ZIP model application: extract static features from basic service attributes and fixed resource configurations, quantify them through the ZIP model, and form a static description of service capabilities; S132. Apply the dynamic induction motor model to monitor service performance indicators and dynamic resource allocation data in real time; by analogy with the operating characteristics of the dynamic induction motor, establish a mathematical model to model the dynamic evolution of service capabilities and capture the changes in service performance over time.

3. The method for intelligent matching of information technology services based on knowledge graph drive according to claim 1 is characterized in that: In S1, the steps of using the multi-head attention mechanism to screen highly relevant entities and form a service network include: S141. Multi-head parallel computing: Through multiple "heads" parallel computing, each "head" captures the relationship between entities from a different perspective; S142. Key information screening: integrating the calculation results of multiple "heads", assigning attention weights to each entity, screening out entities that are highly relevant to service capabilities, removing redundant information, and building a service network that focuses on core business relationships.

4. The method for intelligent matching of information technology services based on knowledge graph drive according to claim 1 is characterized in that: The S3 includes: S311. Similarity calculation: perform similarity calculation on the demand vector and the service feature vector in the initial graph. The similarity calculation is: ； in, The similarity between demand and service; is the demand vector; is the service feature vector; the similarity between the demand and the service is measured by calculating the cosine value of the angle between the two vectors. The closer the value is to 1, the higher the matching degree. S312. A dual DQN network is introduced to dynamically adjust the matching strategy, adaptively adjusting the matching strategy based on similarity calculation results and environmental feedback. Two neural networks with identical structures but different parameter updates are constructed: one for evaluating action value and the other for selecting actions. In the service matching scenario, "action" refers to selecting different service combinations for matching. By continuously interacting with the environment and updating network parameters based on a reward mechanism, the dual DQN network can adaptively adjust the matching strategy according to changing demand, avoiding overestimation. S313. The Adam optimizer generates the objective function. The Adam optimizer adaptively adjusts the learning rate based on the gradient history information of different parameters. During the matching process, a larger learning rate is given to frequently changing parameters to accelerate convergence; a smaller learning rate is given to relatively stable parameters to ensure the accuracy of optimization. The objective function is constructed by comprehensively considering multiple factors such as matching accuracy, service cost, and user satisfaction. The objective function is expressed as: ； Generate objective function value for Adam optimizer; The weight coefficient in the Adam optimizer's objective function is used to adjust the importance of each factor in the objective function. The Adam optimizer adaptively adjusts the learning rate based on the historical gradient information of the objective function for each parameter, and seeks the parameter combination that maximizes the objective function value. Cost of using the service; For user satisfaction; The matching accuracy refers to the degree to which the matching results conform to the user's actual needs; S314. Adversarial sample verification: Design disturbances are made to the original demand data and service data, including adding interference words to the demand text and modifying service feature parameters to generate adversarial samples. The adversarial samples are input into the matching scheme and the changes in the matching results are observed. If the matching scheme can still maintain a high accuracy and stability under the adversarial samples, it means that it has strong robustness. The verified scheme forms a multi-target matching scheme.

5. The method for intelligent matching of information technology services based on knowledge graph drive according to claim 1 is characterized in that: In S4, the step of estimating the difference using the kernel function includes: Based on the service and demand matching results obtained in step S3 and the feedback data after users actually use the service, the kernel function is used to perform difference estimation, and the radial basis function is used to calculate the difference between the predicted value and the actual value. The formula is: ； is the difference between the predicted value and the actual value; is the prediction vector; is the actual feedback vector; It is set according to the fluctuation of historical data; The closer the value is to 0, the greater the difference between the predicted result and the actual situation. When it is less than the set threshold, the graph iteration process is triggered to optimize the service characteristics and matching algorithm.

6. The method for intelligent matching of information technology services based on knowledge graph drive according to claim 1, characterized in that: In S4, the dynamic load modeling step includes: S411. Data Collection and Problem Diagnosis: After confirming significant discrepancies, we initiated in-depth collection of service operation data, covering real-time performance metrics, service call logs, and resource allocation records. Analysis of data from high-concurrency periods revealed CPU utilization consistently exceeding 90%, memory usage consistently exceeding 85%, and significant fluctuations in response time and throughput. Combined with frequent resource request timeouts in service logs, we determined a resource bottleneck was present. S412. Modify the static characteristics of the service based on service performance issues and user feedback. If the service does not meet the promised service level, downgrade the service level attributes. If the service is found to have defects in specific scenarios, add new tags to the fixed resource configuration attributes and provide detailed descriptions to improve the basic attribute description of the service. S413. Use exponential smoothing method to predict service performance indicators. The formula is as follows: ； in, Current cycle predicted response time; The forecast value of the previous period; is the actual measured value of the current cycle; Smoothing coefficient, which reflects the degree of dependence on historical data; The closer it is to 0, the more the model relies on historical forecast values; The closer it is to 1, the more emphasis is placed on the current measured value; The updated response time prediction value, as well as changes in CPU and memory resource usage, are synchronized to the dynamic feature vector of the service in the knowledge graph to ensure that the service characteristics are synchronized with the actual operating status in real time.

7. The method for intelligent matching of information technology services based on knowledge graph drive according to claim 1 is characterized in that: In S4, the specific steps of the closed-loop mechanism to improve matching capabilities are as follows: S421. Entity relationship adjustment: Based on service feature updates and historical optimization experience, analyze the relationship between services and technology nodes. If a technology significantly improves service performance, add a new relationship edge between the corresponding nodes. Based on historical optimization effect evaluation, assign a weight between 0 and 1. A higher weight indicates a closer relationship. At the same time, based on user satisfaction feedback after using the service, the weight of the "usage" relationship between the user node and the service node is dynamically adjusted; when satisfaction decreases, the weight value is proportionally lowered; when satisfaction increases, the weight value is correspondingly increased, so as to more accurately portray the actual relationship between the service and the user; S422. Ontology expansion: extract new demand concepts from user feedback text using natural language processing technology and incorporate them into the ontology structure of the knowledge graph using a bottom-up approach. Based on the semantics and business logic of the new demand concepts, associate them with appropriate ontology subclasses. At the same time, we define relevant attributes for new demand concepts, clarify their characteristics and constraints, and thus expand and enrich the ontology structure of the knowledge graph, so that it can better cover the diverse needs of users and provide comprehensive semantic support for matching; S423. Matching algorithm optimization and dual DQN network parameter adjustment. To make the dual DQN network more adaptable to actual needs, its reward function is adjusted. A penalty term related to service peak performance is added to the original reward function. The new reward function is: ； To balance the similarity parameter; is the cost parameter; is the satisfaction parameter; is the peak performance parameter; It is positively correlated with the degree of response time exceeding the limit during service peak period; When the service response time during peak hours exceeds the preset threshold, The value of increases, resulting in the overall reward value reduce; S424. Adversarial example enhancement: To ensure the robustness of the matching algorithm in extreme scenarios, we generate adversarial examples that simulate extreme loads. These adversarial examples are then fed into the optimized matching model to verify the validity of the matching solutions it outputs. By setting matching accuracy evaluation metrics, we continuously adjust and optimize matching algorithm parameters to ensure that the model can provide reliable matching results in the face of various complex and extreme user needs and service scenarios, thereby improving system stability and reliability. S425. New matching application: When a user submits a service request, the optimized knowledge graph and matching algorithm work together. The knowledge graph, based on the updated entity relationships and ontology structure, selects a set of services that match the user's needs in terms of semantics and technical relevance. The matching algorithm further combines the dynamic characteristics of the services to select the service that best meets the user's needs from the selected set and generates a matching solution. S426. Closed-loop feedback mechanism: new matching results and user feedback will enter S4 again, repeating the difference estimation, service feature update, and algorithm optimization process; Through continuous cycles and iterations, the knowledge graph continuously enriches and corrects service feature information, and the matching algorithm continuously optimizes strategies, ultimately achieving a spiral improvement in the accuracy and adaptability of intelligent matching of information technology services.

Images