Informationization project multi-granularity feature deconstruction and resource adaptation degree evaluation method and system

Abstract

The application discloses an informationization project multi-granularity feature deconstruction and resource adaptation degree evaluation method and system, and belongs to the technical field of information processing and project management. The method comprises the following steps: acquiring a requirement document and preprocessing; extracting features such as project types, functional modules and interface definitions at three granularities of macro, meso and micro; constructing a reference feature library based on historical projects; constructing a multi-granularity feature interaction graph with features as nodes and correlation as edges; generating fused feature embedding by using a graph attention network; outputting personnel, time, equipment and budget demand prediction values by using a multi-head resource prediction network; and comparing with an available resource pool to calculate the adaptation degree score in each dimension. The system corresponds to realize the above modules. The application can deconstruct project features and quantify resource adaptation degrees, and improves the scientificity and efficiency of resource allocation of informationization projects.

Description

Technical Field

[0001] This application relates to the fields of information processing and project management technology, and in particular to a method and system for multi-granularity feature deconstruction and resource suitability assessment of information projects. Background Technology

[0002] As digital transformation continues to advance, various IT projects in government and enterprises (such as ERP implementation, data platform construction, government system integration, customized software delivery, and cloud platform migration) are characterized by requirements being conveyed in text form, diverse technical architecture selections, high coupling of functional modules, and significant differences in construction cycles and resource investment scale. During the project initiation or launch phase, organizations often need to answer the following core questions: Based on the current project requirements description, what are the approximate manpower (person-days / person-months), timeframe, computing power / equipment, and budget required? And can the existing available resource pool support the project? These questions determine the project initiation, scheduling, and resource allocation strategies.

[0003] To address these needs, the industry has long developed several relatively mature assessment approaches, but each has its own structural shortcomings. Firstly, qualitative or semi-quantitative assessments, such as the expert judgment method (e.g., the Delphi method), rely on experienced project managers or technical consultants to manually interpret and "count" requirements documents based on their experience with similar projects. The advantage is that it can incorporate difficult-to-quantify contextual factors when information is incomplete. The disadvantages are strong subjectivity, significant differences in conclusions drawn by different assessors for the same material, and difficulty in establishing reusable and sustainably calibrated data loops from the assessment results. Secondly, parametric models, such as COCOMO / COCOMOII and Function Point Analysis (FPA / IFPUG / NESMA), express workload / cost as a mathematical form of "scale × cost driver factors." While these models offer interpretability and generalizability in engineering management, their input still relies heavily on lines of code or manually counted / calculated function scale. In the early stages of a project, they often face challenges such as "difficulty in determining scale, complex counting rules, and the need for local historical baseline calibration of parameters." Furthermore, the cost driver factor system of these models is still biased towards macro-level statistics and lacks the ability to absorb structural information at the meso-micro level, such as "which specific technology stacks were used, how many external systems need to be integrated, the data architecture, and the density of business rules." Third, analogical estimation methods, represented by tools like ANGEL, represent historical projects as vectors in a multi-dimensional feature space. They retrieve nearest neighbor projects using similarity metrics (such as Euclidean distance) and use their actual consumption as the estimation benchmark for the current project. This approach brings the evaluation closer to data-driven methods, but it still faces two common problems in engineering: First, the feature space is usually a "flat vector," and the historical project database mainly stores project-level summary attributes (type, scale range, team capabilities, etc.), losing the co-occurrence structure, dependency amplification effect, and semantic coupling relationship between features. Second, similarity matching can only provide a "looks like" but cannot explain "why the internal structure drives resources." When requirements change locally (such as adding a new type of external system to connect to or switching data architecture selection), the analogy results are prone to mismatches where "the surface is similar but the resource-driving logic is different."

[0004] On the other hand, recent studies have attempted to introduce graph neural networks (GNN / GAT) to improve the prediction of project time and cost. However, their modeling objects are generally project scheduling networks (activity-on-node / activity-resourcegraph) composed of "activity nodes + prerequisite dependencies / resource allocation," that is, capturing the propagation effect between tasks and resource contention through known work breakdown structures or scheduling topologies. In these schemes, the node semantics of the graph come from "pre-decomposed WBS / activity list / resource ledger." Therefore, existing "graph-based" solutions do not solve the problem of "how to generate this graph from the requirements text," while existing "text feature extraction" solutions often stop at the keyword / classification label level, ultimately flattening the results into independent variables and feeding them into traditional regression or tree models. The nonlinear interactions between features can only be approximated by manually designing cross terms.

[0005] In summary, the following core contradictions remain unresolved in the field of information technology project resource demand assessment: the project information carried by the demand document is multi-layered and features are coupled, but the existing assessment system lacks a complete link from "demand text → automatic multi-granular feature deconstruction → explicit graphing of structural relationships between features → differentiable graph representation learning and end-to-end multi-dimensional resource prediction → resource pool fit feedback"; the consequence is that the assessment can only degenerate into empirical analogy or flat parameter model, resulting in: (1) structural information that has key amplification on resource consumption at the meso and micro levels (technology stack selection, integration interface form, data entity complexity, business rule density, etc.) cannot stably and automatically participate in the calculation; (2) co-occurrence associations and cross-granular semantic coupling between features (e.g., the impact of architecture selection on interface scale and testing manpower) are linearized or ignored; (3) the matching from historical projects to current projects lacks structured interpretability, and it is difficult to transmit the impact of local demand changes in a fine-grained manner; (4) manpower / time / equipment / Budget forecasts are fragmented and lack a unified adaptation conclusion linked to the available resource pool. The final output remains at "giving a number" rather than "giving an executable resource matching judgment and risk warning." Therefore, a solution is needed that can use multi-granular features themselves as graph nodes, use intra-level co-occurrence associations and cross-level semantic associations as edges, and complete project embedding learning on the feature interaction graph through a graph attention mechanism, thereby systematically improving the accuracy of resource prediction and the quality of adaptation decisions. Summary of the Invention

[0006] To address the shortcomings of existing information technology project resource demand assessments, such as the inability to structurally extract the characteristic information carried in project requirement documents into computable inputs, the flattening of coupling relationships between features, and the difficulty in forming a closed-loop adaptation judgment between resource prediction results and available resource pools, this application aims to provide a method and system for multi-granularity feature deconstruction and resource suitability assessment of information technology projects. This method can improve the structural traceability, predictive rationality, and operability of resource allocation decisions while reducing human intervention and subjective bias.

[0007] To achieve the above objectives, this application adopts the following technical solution:

[0008] This application provides a method for multi-granularity feature deconstruction and resource suitability assessment of information technology projects, including:

[0009] Step S1: Obtain the requirements document of the project to be evaluated, and obtain standardized requirements text through text preprocessing;

[0010] Step S2 involves multi-granularity feature deconstruction of the standardized requirement text: At the macro-granularity, project type, business domain, application scenario, and organizational scale are extracted through text classification and domain dictionary matching to form a macro-feature set; at the meso-granularity, functional modules, technology stack selection, data architecture type, and integration interface type are extracted through dependency parsing and template matching to form a meso-feature set; at the micro-granularity, interface definitions, data entities, functional atoms, and business rules are extracted through named entity recognition and rule reasoning to form a micro-feature set.

[0011] Step S3: Construct a reference project feature library based on historical project data. Each record contains multi-granularity feature vectors of completed projects and their actual resource consumption data, and stores them in partitions according to project type.

[0012] Step S4: Using all the features obtained in step S2 as nodes and the relationships between features as edges, construct a multi-granularity feature interaction graph; wherein, when the co-occurrence frequency of two features in the reference library within the same granularity exceeds a preset threshold, establish an intra-level association edge, and the weight is determined by the co-occurrence frequency and mutual information; when two features at different granularities have semantic relationships, establish a cross-level association edge, and the weight is determined by the semantic relationship strength and conditional probability.

[0013] Step S5: Input the interaction graph into the pre-trained graph attention network, and after multiple layers of multi-head attention convolution and global attention pooling, obtain the fusion feature embedding vector representing the entire project.

[0014] Step S6: Embed the fused features into the pre-trained multi-head resource prediction network, and output the demand forecasts for personnel, time, equipment and budget resources through a shared layer and four parallel prediction heads, respectively.

[0015] Step S7: Compare the predicted resource demand with the available resource pool, calculate the fit score and resource gap for each dimension, and then weight them according to the preset weights to obtain the comprehensive fit score, thereby generating a resource fit score assessment report.

[0016] As a preferred technical solution, step S2, the extraction process of macro features includes: inputting the standardized requirement text into a pre-trained text classification model to obtain classification results for project type and business domain; using a preset domain dictionary to perform keyword matching on the standardized requirement text to identify application scenario terms and organizational scale indicators; merging the text classification results and dictionary matching results to remove duplicates, obtaining a macro feature set, where each macro feature contains feature terms and feature confidence, and the feature confidence is determined by the similarity between the prediction results of the text classification model and the dictionary matching results.

[0017] As a preferred technical solution, step S2, the extraction process of meso-level features includes: performing dependency parsing on the standardized requirement text to obtain the dependency relationship tree of each sentence; matching the dependency relationship tree using a predefined functional module template library to identify functional module features; identifying the technology stack selection features of the standardized requirement text using a predefined technology stack terminology library and regular expression patterns; identifying the data architecture type features of the standardized requirement text using a predefined data architecture pattern; identifying the integration interface type features of the standardized requirement text using a predefined interface type template; and merging the identification results of the above four types of features to obtain the meso-level feature set.

[0018] As a preferred technical solution, step S3, the construction process of the reference project feature library includes: for each completed project, extracting its multi-granularity feature vector and actual resource consumption vector; cleaning the feature vectors and resource consumption vectors of all completed projects, removing records with missing or obvious abnormalities in the resource consumption data; dividing the cleaned data into several partitions according to project type, with each partition arranged in order of project completion time; merging partitions with fewer than the preset requirement into other types of partitions; the reference project feature library supports incremental updates after adding new completed project data.

[0019] As a preferred technical solution, in step S4, the edge weights are determined as follows: For intra-level related edges, the edge weights are determined by weighting the normalized standard mutual information and the normalized co-occurrence frequency. The normalized standard mutual information is calculated based on the probability of the feature appearing in the reference item feature library and the probability of co-occurrence. The normalized co-occurrence frequency is calculated based on the ratio of the number of items where the feature appears simultaneously to the total number of items. The two are combined using a preset balance coefficient. For cross-level related edges, the edge weights are determined by weighting the semantic similarity based on pre-trained word vectors and the conditional probability. The conditional probability refers to the probability that another feature also appears given that one feature appears. The two are combined using a preset balance coefficient.

[0020] As a preferred technical solution, step S5, the training process of the graph attention network model includes: using completed projects in the reference project feature library as training samples, constructing a multi-granularity feature interaction graph for each training sample as model input, and using the actual resource consumption vector of the project as supervision label; jointly training the graph attention network model and the multi-head resource prediction network end-to-end, using a weighted combination loss function as the training objective; using an optimizer for gradient backpropagation and employing a learning rate decay strategy; setting the number of training rounds and batch size; employing an early stopping strategy during training, stopping training when the validation set loss no longer decreases within several consecutive rounds, and saving the model parameters that perform best on the validation set; and dividing the dataset into a training set, a validation set, and a test set according to a preset ratio.

[0021] As a preferred technical solution, in step S5, the node embeddings are updated through a multi-head attention mechanism in the first and second graph attention convolutional layers: for each attention head, the normalized attention coefficient of the neighboring nodes to the current node is calculated, and the features of the neighboring nodes are weighted and aggregated using this coefficient. After processing by the activation function, the outputs of all attention heads are concatenated to obtain the new node embedding of this layer; the normalized attention coefficient is obtained by concatenating the query vector and the key vector and then processing it with the activation function and normalization; both the first and second graph attention convolutional layers have a preset number of attention heads and corresponding output dimensions, and the negative slope parameter of the activation function is set as needed; the global attention pooling layer performs a weighted summation of all node embeddings output by the second layer through learnable global attention weights, and concatenates the weighted summation result with the max pooling result to obtain the fused feature embedding vector.

[0022] As a preferred technical solution, in step S6, the training loss function of the multi-head resource prediction network is a composite loss function, which is composed of a weighted combination of the mean squared error and the mean absolute percentage error of each resource dimension, and an L2 regularization term is added to the model parameters to prevent overfitting; the loss weights of each resource dimension are preset according to the importance of the resource, and the global error term and regularization coefficient are set empirically; the shared fully connected layer contains multiple fully connected networks, each followed by an activation function and a random dropout layer to prevent overfitting, and finally outputs a fixed-dimensional shared feature representation; each of the four resource prediction heads independently contains multiple fully connected networks, each followed by an activation function, and the last layer has no activation function, directly outputting the single demand prediction of the corresponding resource.

[0023] As a preferred technical solution, step S7, the generation of the resource suitability assessment report also includes: classifying the resource status of each dimension into three levels—sufficient resources, critical resources, and insufficient resources—based on the positive and negative direction and magnitude of the resource gap in each dimension, with each level determined according to a preset warning threshold; classifying the comprehensive suitability level into multiple preset levels based on the numerical range of the comprehensive suitability; and marking risk warning information and generating resource allocation suggestions in the assessment report for resource dimensions determined to be insufficient or projects with low comprehensive suitability levels.

[0024] This application also provides a system for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects, including:

[0025] The text preprocessing module is used to obtain the requirements document of the project to be evaluated and perform text preprocessing to obtain standardized requirements text.

[0026] The multi-granularity feature deconstruction module is used to extract project features from standardized requirement texts at three granularity levels: macro, meso, and micro, to obtain macro feature sets, meso feature sets, and micro feature sets.

[0027] The reference project feature library module is used to store multi-granularity feature vectors of historical completed projects and corresponding actual resource consumption data, and supports partitioned retrieval and incremental updates by project type.

[0028] The feature interaction graph construction module is used to construct a multi-granularity feature interaction graph with multi-granularity features as nodes and inter-feature relationships as edges, including the establishment of intra-level and cross-level relationship edges and the calculation of edge weights.

[0029] The graph attention network inference module is used to load pre-trained graph attention network models, perform multi-head graph attention convolution and global attention pooling on multi-granularity feature interaction graphs, and output fused feature embedding vectors.

[0030] The multi-head resource prediction module is used to load a pre-trained multi-head resource prediction network, embedding the fused features into a vector through a shared fully connected layer and four parallel resource prediction heads, and outputting the demand predictions for personnel, time, equipment and budget resources respectively.

[0031] The resource suitability assessment module is used to calculate the suitability between the predicted resource demand and the available resource pool, obtain the suitability scores and resource gaps for each dimension, and generate a comprehensive suitability level according to preset weights, thus forming a resource suitability assessment report.

[0032] Compared with the prior art, the beneficial effects of this application are as follows:

[0033] Current IT project resource assessments often rely on qualitative judgments based on expert experience or analogies with the flat features of historical projects. Feature extraction is often limited to macro-level labels such as project type and scale, and each feature is treated as an independent variable. This results in the discarding of a large number of meso-level structures and micro-level details in the requirements document that affect resource consumption. The co-occurrence amplification effect and cross-granularity semantic coupling between features cannot be incorporated into the calculation process. The prediction results are subjective, coarse-grained, and difficult to interpret, and cannot be directly matched with the organization's available resource pool to form an actionable conclusion. This application first standardizes and preprocesses the project requirements document, then extracts features at three granular levels—macro, meso, and micro—as well as features related to type / domain / scenario / scale, functional modules / technology stack / data architecture / integration interfaces, and interface definitions / data entities / functional atoms / business rules. This makes the "structural driving forces" buried in the original unstructured text explicitly emerge in the form of an enumerable set of features, thereby providing a consistent and traceable input basis for subsequent quantitative modeling, reducing reliance on the personal experience of evaluators, and improving the stability and reproducibility of evaluation results across different projects and different evaluators.

[0034] Furthermore, instead of performing simple regression using flat feature vectors, this application constructs a feature interaction graph using multi-granular features as nodes and intra-level co-occurrence associations and cross-level semantic associations as edges. A pre-trained graph attention network is then used to perform multi-head attention convolution and global pooling on this graph to obtain item-level fusion embeddings. This allows the information transmission of "who frequently coexists with whom and who has a semantic driving relationship with whom" between features to participate in the information transmission in the form of graph structure weights. This naturally accommodates nonlinear interactions and coupling amplification effects at the representation level, improving the systematic bias that cannot be captured when relying solely on manually setting cross terms or global statistical means. At the same time, the graph structure itself provides verifiable association evidence, making it easier to locate the affected feature subgraphs and associated edges when local changes occur in requirements, making resource impact assessment more granular and interpretable. Subsequently, the system integrates and embeds a multi-head resource forecasting network to output four types of demand forecasts: manpower, schedule, equipment, and budget. This avoids inconsistencies between dimensions caused by isolated estimations of single resources. Finally, the forecasts are compared with the available resource pool to calculate the dimensional fit and resource gap, and then weighted to form a comprehensive fit level. The assessment report provides the sufficiency / critical / insufficiency status of each dimension, along with risk warnings and allocation suggestions. This elevates resource assessment from "giving a rough estimate" to an actionable conclusion for resource management, directly supporting project prioritization, scheduling window selection, and coordinated allocation decisions for personnel, external procurement, and budget. Overall, the solution can continuously absorb new completed project samples by relying on a reference project feature library partitioned by project type and supporting incremental updates. This allows the assessment capability to continuously self-calibrate as the organization's project data accumulates, balancing automation, assessment accuracy, and engineering feasibility. Attached Figure Description

[0035] Figure 1 A flowchart illustrating the steps of a multi-granularity feature deconstruction and resource suitability assessment method for information technology projects;

[0036] Figure 2 A flowchart outlining the steps to obtain standardized requirements text through text preprocessing for the project to be evaluated.

[0037] Figure 3 A flowchart illustrating the steps involved in multi-granularity feature deconstruction of standardized requirement text;

[0038] Figure 4 A flowchart outlining the steps involved in building a reference project feature library based on historical project data.

[0039] Figure 5 This is a multi-granularity feature interaction graph;

[0040] Figure 6 This is a schematic diagram of the graph attention network model.

[0041] Figure 7This is a schematic diagram of the structure of an information technology project multi-granularity feature deconstruction and resource suitability assessment system. Detailed Implementation

[0042] To enable those skilled in the art to better understand the present application, the technical solutions in specific embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0043] This application provides a method for multi-granularity feature deconstruction and resource suitability assessment of information technology projects. The method takes the requirements document of the information technology project as input, and goes through text standardization preprocessing, three-layer granularity feature automatic deconstruction, reference feature library retrieval, feature interaction graph construction, graph attention inference, multi-dimensional resource prediction, and resource pool suitability assessment. Finally, it outputs a resource suitability assessment report that includes suitability scores for each dimension, resource gap amount, and comprehensive suitability level.

[0044] like Figure 1 As shown below, the explanation will be carried out step by step.

[0045] Step S1: Obtain the requirements document of the project to be evaluated, and obtain standardized requirements text through text preprocessing.

[0046] Specifically, such as Figure 2 As shown, in actual execution, step S1 first reads the requirements document of the information technology project to be evaluated from the project management system. The requirements document can be in the format of a Word document, PDF document, Markdown file, or plain text survey summary. The system extracts text from the requirements document of the project to be evaluated, converting non-plain text formats into continuous character sequences. Then, a text preprocessing pipeline is executed, including the following sub-steps:

[0047] The first sub-step S1 01 removes irrelevant characters. Based on a predefined set of rules for irrelevant characters, the system filters out page number markers, headers and footers, table of contents numbers, special control characters, garbled fragments, and decorative symbols that do not participate in semantic expression, while retaining Chinese characters, English letters, numbers, common punctuation marks, and key symbols in the project domain.

[0048] The second sub-step, S1 02, involves sentence segmentation. Following Chinese punctuation rules, the system divides the entire document into sentence sequences using periods, question marks, exclamation marks, semicolons, and line breaks as sentence boundaries. For each sentence, its original sentence number is retained for later retrieval during cross-sentence dependency analysis.

[0049] The third sub-step S1 03 is word segmentation. The system loads a Chinese word segmentation dictionary for the field of information technology projects. This dictionary pre-includes common IT terms, software product names, technical framework names, system module names, interface protocol names, and abbreviations of departments and organizations. It uses a word segmentation strategy that combines dictionary-based maximum matching with statistical language models to segment each sentence and output a word sequence.

[0050] The fourth sub-step S1 04 is to remove stop words. The system reads a predefined stop word list, which contains meaningless function words, conjunctions, auxiliary words, modal particles, and generalized words that frequently appear in the requirements document but do not carry project feature information. The system removes the lexical units that match the stop words in the word segmentation results and retains the lexical units that have domain semantic load.

[0051] After the above preprocessing, the system reorganizes all retained lexical units into standardized requirement text. This standardized requirement text can be stored as a cleaned list of sentences, with each sentence associated with its lexical sequence and part-of-speech tagging sequence, for direct use in the subsequent three-layer feature deconstruction step S2. The beneficial effect of this step is that it transforms the messy requirement documents into a structurally consistent text representation through a unified preprocessing pipeline, reducing noise interference in the subsequent feature extraction stage and providing a clean input foundation for macro-classification, dependency parsing, and entity recognition.

[0052] Step S2 involves deconstructing the standardized requirement text using multi-granularity features.

[0053] At the macro-granularity level, project type, business domain, application scenario, and organizational scale are extracted through text classification and domain dictionary matching, forming a macro-feature set. At the meso-granularity level, functional modules, technology stack selection, data architecture type, and integration interface type are extracted through dependency parsing and template matching, forming a meso-feature set. At the micro-granularity level, interface definitions, data entities, functional atoms, and business rules are extracted through named entity recognition and rule reasoning, forming a micro-feature set.

[0054] like Figure 3 As shown below, the specific implementation details of the extraction process at the three granularity levels are given respectively.

[0055] a. Specific implementation at the macro-granular level.

[0056] The system inputs standardized requirement text into a pre-trained text classification model to obtain project type classification results and business domain classification results. The pre-trained text classification model can employ a pre-trained language model based on a Transformer encoder architecture, such as BERT or its variants. The model's output layer interfaces with two independent multi-class classification heads. The first classification head outputs the probability distribution of project types, including ERP implementation, data platform construction, government system integration, customized software delivery, cloud platform migration, and other custom types. The second classification head outputs the probability distribution of business domains, including finance, manufacturing, government affairs, healthcare, education, retail, and logistics. The system selects the category with the highest probability value as the corresponding classification result, while retaining the probability value as an initial confidence score.

[0057] Meanwhile, the system utilizes a pre-defined domain dictionary to perform keyword matching on standardized requirement text, identifying application scenario terms and organizational scale indicators. The domain dictionary contains three categories of terms: the first category consists of application scenario trigger words, such as "approval workflow, expense reimbursement, warehousing, outbound, reconciliation, reports, portal, single sign-on, mobile terminal," etc., which are matched using a combination of word-level exact matching and longest prefix matching. The second category consists of organizational scale indicators, such as "thousands of employees, tens of thousands of employees, more than N branches, number of outlets," etc., which are extracted from the text using regular expressions to extract numerical descriptions related to personnel or organizational scale. The third category consists of supplementary business domain terms, used for secondary validation and recall beyond text classification.

[0058] The system merges and deduplicates text classification results with dictionary matching results to obtain a macro feature set. Each macro feature in the macro feature set is represented as a tuple containing the feature term and feature confidence. Feature confidence is determined by the combined similarity between the prediction results of the text classification model and the dictionary matching results. Specifically, if a term is matched by both the classification model and the dictionary, its confidence is a weighted combination of the classification probability and the dictionary similarity score. If it is matched only by the classification model, the classification probability is used as the confidence. If it is matched only by the dictionary, the normalized matching similarity is used as the confidence. The merging and deduplication process involves retaining the term with the higher confidence when different expressions of the same concept point to the same normalized term. The final macro feature set covers four aspects: project type, business domain, application scenario term list, and organizational scale indicators.

[0059] The principle behind this part is that pure text classification can infer the item's affiliation from global semantics but may miss local keywords, while pure dictionary matching can accurately locate terms but cannot handle unseen words and complex expressions. The fusion and complementarity of the two make the macro features have both global discriminative power and retain local interpretability.

[0060] b. Specific implementation of the meso-level granularity.

[0061] The system performs dependency parsing on standardized requirement texts to obtain dependency trees for each sentence. The dependency parser can employ a deep learning-based dependency parsing model to output the parent node number, dependency label, and dependency arc direction for each word in each sentence. Dependency labeling includes subject-verb, verb-object, attributive-head, prepositional phrase, and coordinate relationships.

[0062] The system uses a predefined functional module template library to match dependency trees and identify functional module features. This template library is compiled from historical data of IT projects. Each template describes the syntactic form of a functional module, such as a verb-object structure with a core verb and object, modified by a specific attributive clause. The system scans each sentence using subtree pattern matching on the dependency tree. When the syntactic structure of a subtree matches the template and the key content words fall within the functional module dictionary, the module name is recorded as a functional module item in the meso-feature set.

[0063] The system utilizes a predefined technology stack terminology database and regular expression patterns to identify technology stack selection features in standardized requirement texts. The terminology database includes common front-end and back-end framework names, middleware names, database product names, programming language names, and deployment environment types. The system scans the terminology database using segmented word sequences, performing forward maximum matching and substring matching, and combining this with modifiers in the context window to determine whether the technology is used in the development stack or the runtime stack. Matched technology stack items are recorded as technology stack selection items in the meso-feature set.

[0064] The system uses predefined data architecture patterns to identify the data architecture type characteristics of standardized requirement texts. The data architecture patterns include text descriptions of patterns such as "centralized relational database, read-write separation, database sharding, data warehouse layering, data lake, real-time stream processing, and master data management". The system determines which data architecture form or which combination of data architectures is described in the text by combining keywords and syntactic constraints.

[0065] The system utilizes predefined interface type templates to identify the integration interface type characteristics of standardized requirement texts. These templates include descriptions of integration methods such as "RESTAPI, SOAPWebService, message queue, file transfer, direct database connection, and SDK embedding." The system retrieves trigger verbs and their object phrases (e.g., docking, integration, calling, synchronization, subscription, push) from the dependency tree and determines the interface type based on the protocol terms and modifying structures within the object phrases.

[0066] The identification results of the above four types of features are combined to obtain the meso-level feature set. The beneficial effect of this step is that the meso-level granularity captures the structural factors that truly drive the differences in workload in the project, namely the complexity of module combination, the diversity of technology stack, the maturity requirements of data architecture, and the size of external integration surface. This information cannot be expressed by macro-level type labels.

[0067] c. Specific implementation of the micro-granularity level.

[0068] The system extracts features at the interface definition, data entity, and functional atomic levels through named entity recognition. The named entity recognition model can employ a sequence labeling architecture, defining tags such as interface entities, data table entities, field entities, business object entities, and rule condition entities. After the model annotates the standardized requirement text sentence by sentence, the system extracts specific interface paths or interface names, data entity names, and field names from the annotation span, and extracts their attribute information from the context using rules, such as the interface's request / response format, primary and foreign key implications of data entities, and the type range of fields.

[0069] The system extracts business rule features through rule-based reasoning. The rule reasoning engine scans conditional clauses and sentence fragments containing keywords such as "if, then, otherwise, should, prohibit, allow, must" in the dependency tree, extracting conditional items, action items, and constraint objects, and structuring them into business rule entries stored in the micro-feature set. Each business rule entry includes a summary of the rule triggering conditions, references to the data entities or functional modules involved, and a rule type classification.

[0070] Ultimately, the micro-feature set, together with the macro-feature set and the meso-feature set, constitutes a complete multi-layered feature profile of the project to be evaluated. The three sets are not interchangeable and answer three progressively deeper questions: What is this project? What meso-level components does it consist of? What are its specific interfaces, data, and rule details?

[0071] Step S3: Construct a reference project feature library based on historical project data.

[0072] Each record contains multi-granularity feature vectors of completed projects and their corresponding actual resource consumption data, and is stored in partitions according to project type.

[0073] Specifically, before implementing the method described in this application, the organization must first complete the initial construction of the reference project feature library. For each historically completed IT project, the system retrieves the project's requirements document from that year and processes it using the same preprocessing and three-layer deconstruction pipeline as steps S1 and S2 to obtain the project's macro-feature set, meso-feature set, and micro-feature set. Simultaneously, the system reads the project's actual resource consumption data from the project performance management system, including actual personnel consumption, actual time consumption, and actual budget consumption; equipment consumption can be summarized from asset requisition records.

[0074] The system converts the feature sets of all completed projects into multi-granularity feature vector representations. One implementation involves maintaining a global feature vocabulary, constructed from feature terms of all historical projects during feature library initialization, with each feature term mapped to a dimension index. Macro- and meso-feature sets are populated into the corresponding dimensions of the vectors using a bag-of-words approach with confidence weighting. Micro-feature sets, due to their variable number and finer granularity, can participate in vectorization using aggregated statistics, such as the number of micro-features, interface definitions, data entities, and business rules, as additional dimensions. Thus, each record forms a pair of "feature vector plus actual resource consumption vector".

[0075] The system performs data cleaning on all records, removing abnormal records where any dimension of resource consumption data is empty or obviously unreasonable. One method of anomaly detection is to calculate the mean and standard deviation for each resource dimension, and to remove records that deviate from the mean by more than three times the standard deviation. At the same time, the system checks for logical contradictions where the resource consumption values ​​are seriously mismatched with the project scale.

[0076] After cleaning, the data is divided into several partitions according to project type. Within each partition, projects are arranged in chronological or descending order of completion time, forming an ordered reference sequence for similar projects on the timeline. This facilitates the selection of time decay weighting for subsequent co-occurrence statistics. For rare types with fewer projects than the minimum sample size threshold within a partition, the system merges them into the semantically closest parent type partition to avoid sparsity issues.

[0077] The reference project feature library supports incremental updates after adding completed project data, eliminating the need for a full rebuild. Whenever a new project is completed and its actual consumption is confirmed and archived, the system performs steps S1 and S2 of the deconstruction process on its requirements document to obtain a new feature set, appends it to the corresponding type partition, and updates the co-occurrence statistics table and feature probability table within that partition. In this way, as the organization accumulates projects, the coverage and statistical reliability of the feature library gradually improve, and the evaluation accuracy self-calibrates accordingly.

[0078] Step S4: Using all the features obtained in step S2 as nodes and the relationships between features as edges, construct a multi-granularity feature interaction graph, such as... Figure 5 As shown.

[0079] Specifically, when the co-occurrence frequency of two features within the same granularity level exceeds a preset threshold in the reference library, an intra-level association edge is established, with the edge weight determined by the co-occurrence frequency and mutual information. When two features at different granularity levels have semantic association, a cross-level association edge is established, with the edge weight determined by the semantic association strength and conditional probability.

[0080] Specifically, the system collects all feature terms from the macro-feature set, meso-feature set, and micro-feature set output in step S2, and creates a graph node for each unique term. All nodes constitute a node set V. Each node carries node attributes, including at least: node ID, original text of the feature term, label of the granularity level, and source confidence.

[0081] The next step is to build edges in two categories.

[0082] The first type is intra-level associative edges. The system iterates through each granularity level, pairing nodes together, and for each pair of nodes... and Query the co-occurrence statistics in the feature library of reference projects, that is, count how many historical projects simultaneously exhibit the same feature in the same project. and characteristics Calculation features Probability of appearance in the library Equal to appear The number of projects divided by the total number of projects. Similarly, the probability of simultaneous occurrence Equivalent to appearing simultaneously and The number of projects divided by the total number of projects. The normalized standard mutual information is calculated as follows: The value falls within arrive Between, the larger the value, the more it indicates and The co-occurrence of these factors cannot be independently explained by their respective marginal probabilities. Normalized co-occurrence frequency. .when When it exceeds the preset first threshold, and Establish undirected edges between them, with edge weights. , To set the balance coefficient, we can take... In this way, frequently co-occurring feature pairs with strong mutual information are assigned heavier edge weights, reflecting their structural binding relationship in terms of resource consumption, such as the coupling between microservice architecture and the proliferation of message queue interfaces.

[0083] The second type is cross-level association edges. The system associates nodes belonging to different granularity levels. and Calculate the semantic association strength. Semantic association strength is calculated based on pre-trained word vectors. Feature terms or their associated phrases are fed into a pre-trained word vector model to obtain vector representations, and then the cosine similarity is calculated as the semantic association strength. Simultaneously calculate the conditional probability. .when Greater than the semantic threshold or When the conditional probability threshold is greater than 1, and Establish cross-level association edges, edge weights , To set the balance coefficient, we can take... Cross-level edges embody, for example, the semantic traction from the macro-level government system integration type to the meso-level interface type equal to SOAP, and the entity association from the meso-level technology stack including Redis to the micro-level cache-related interfaces.

[0084] Completed diagram notation ,in This is the union of the two types of edges mentioned above. The advantage of a graph structure is that subsequent graph neural networks can directly propagate information at nodes and edges, instead of flattening features into independent variables. This step is the key transition point for the entire scheme from traditional feature vector methods to structured representations.

[0085] Step S5: Input the interaction graph into the pre-trained graph attention network, and after multi-layer multi-head attention convolution and global attention pooling, obtain the fusion feature embedding vector representing the entire project.

[0086] like Figure 6 As shown, the graph attention network model includes a first graph attention convolutional layer, a second graph attention convolutional layer, and a global attention pooling layer.

[0087] During implementation, the system first initializes the node feature vector for each node in the graph. The initialization method can use a pre-trained word vector lookup table of terms as the basic dimension, and then concatenate the node's confidence scalar, granular hierarchical one-hot encoding, and degree centrality as supplementary dimensions to make different nodes comparable in the same coordinate space.

[0088] The first image shows an attention convolutional layer that uses... Each attention head performs a multi-head attention convolution on the input node features. Learnable weight matrix The input node features are linearly transformed to a high-dimensional space. For the nodes... and its neighboring nodes Calculate the attention coefficient ,in For the first The attention vectors of each attention head are used. Normalized attention coefficients are obtained after Softmax normalization. .node In the The update under the size is represented as . The outputs of each node are concatenated along the feature dimension to obtain the first-layer output node embedding. A typical configuration is as follows: Each head outputs dimensions After splicing dimension.

[0089] The second attention convolutional layer takes the output of the first layer as input and employs... Each attention head undergoes further convolution, with the structure similar to the first layer. A typical configuration is as follows: Each head outputs dimensions After splicing Dimensionality. The two-layer stacking enables nodes to perceive the characteristic interaction structure within a two-hop range.

[0090] The global attention pooling layer performs weighted sum pooling and max pooling operations on all node embeddings from the second layer's output, then concatenates the results. In the weighted sum pooling part, each node... Global attention weights ,in and For learnable parameters, For the nodes output by the second layer Embedded. The weighted summation result is: Max pooling takes the maximum value of all node embeddings in each dimension. The two are then concatenated to obtain the fused feature embedding vector. Its dimensions are fixed and decoupled from the number of nodes in the graph, so it can be fed into a subsequent fixed-width fully connected network.

[0091] The training process of the graph attention network model is as follows: The system uses completed items from the reference item feature library as training samples. For each training sample, step S4 is performed to construct the input graph according to its multi-granularity feature set. The actual resource consumption vector of the item is used as the supervision label. The graph attention network and the multi-head resource prediction network in step S6 are jointly trained end-to-end. The loss function adopts a weighted combination of mean squared error (MSE) and mean absolute percentage error (MAPE), and an L2 regularization term is added. The optimizer is Adam, and the initial learning rate is set to... A cosine annealing scheduling strategy is employed to make the minimum learning rate approach the minimum value. The number of training epochs is set to... Batch size (BatchSize) is set to To prevent overfitting, an early stopping strategy is used to monitor the validation set loss. If continuous... If the parameters do not decrease, training stops and the system rolls back to optimal parameters. Data is then processed according to... The dataset is divided into training, validation, and test sets. Since the graph construction process is a deterministic pipeline operation for each sample, training can be performed in batches. Graph batching adopts a strategy of filling in nodes or subgraph sampling.

[0092] The core principle of this step is that GAT's attention mechanism does not require manual specification of which feature combinations affect resources. Instead, it learns through data-driven methods which edges on the feature interaction graph contribute more to the propagation of the final resource consumption, thus implicitly modeling nonlinear interactions, which is something that traditional linear regression or decision trees cannot do.

[0093] Step S6: Input the fused feature embedding vector into the pre-trained multi-head resource prediction network.

[0094] The system outputs demand forecasts for personnel, time, equipment, and budget resources by sharing a fully connected layer and four parallel resource forecast heads.

[0095] The structure of the multi-head resource prediction network is as follows. The shared fully connected layer contains two sequentially connected fully connected layers. The first layer will... From the fusion dimension (e.g.) spliced (or the aforementioned actual splicing dimension) mapped to Dimension, activation function is , followed by (ratio The second layer from Dimension mapping to Dimension, activation function is , followed by (ratio ). Output Dimensional vectors are the shared feature representations. It encodes the structural information of the entire project and has been dimensionality reduced and noise removed.

[0096] The four parallel forecast heads are personnel resource forecast head, time resource forecast head, equipment resource forecast head, and budget resource forecast head. Each head is independent but shares the same... Input. Each prediction head contains three fully connected layers. Taking a person's head as an example, the first layer... activation The second layer activation The third layer Directly outputs a scalar without an activation function The other three head structures are symmetrical and output respectively. , , .

[0097] The loss function during the training phase is .in . Desirable , Pick , Pick , Pick , Desirable , Desirable This composite loss takes into account both absolute and relative errors, weights different resource dimensions according to their importance, and applies regularization constraints to the parameter size to prevent overfitting.

[0098] During the inference phase, the h_fused from step S5 is fed into the pre-trained network with fixed parameters to obtain the four-dimensional prediction in one go. The beneficial effect of this structure is that the four resource dimensions share the underlying item structure representation to avoid inconsistencies caused by learning from scratch, while retaining the differences in the output scale of each dimension through independent prediction heads.

[0099] Step S7: Compare the predicted resource demand with the available resource pool, calculate the fit score and resource gap for each dimension, and then weight them according to the preset weights to obtain the comprehensive fit score, thereby generating a resource fit score assessment report.

[0100] Specifically, the system reads the vector of currently available resource pools from the resource management system. Includes available personnel stock Available time window conversion Available equipment conversion quantity and available budget For each dimension Calculate the fit The closer the value is to This indicates that the closer the supply and demand match, the closer they are. This indicates a larger gap. Resource disparity is also calculated. Positive numbers indicate insufficient quantity, while negative numbers indicate surplus quantity.

[0101] according to The state is divided into three levels based on the sign, direction, and magnitude. Taking the personnel dimension as an example, a preset warning threshold is established. ,when When resources are deemed sufficient, Determine resource criticality when Resource insufficiency is determined at that time. Each dimension—time, equipment, and budget—has its corresponding threshold. , , .

[0102] Overall compatibility Equal to each dimension take Summation, Use the same definition as the loss function to ensure consistency in the evaluation criteria. Divide the levels according to numerical ranges. For high compatibility, Medium fit For low-level adaptation, Not a good match.

[0103] For any dimension deemed to have insufficient resources or a project with an overall fit below medium fit, the system marks a risk warning in the assessment report and generates resource allocation suggestions, such as suggesting delaying the schedule, borrowing manpower, splitting the delivery into phases, increasing the budget, or replacing high-consumption technology stacks.

[0104] The final resource suitability assessment report should include at least the following: project identifier, four-dimensional resource prediction, available resource pool reading, suitability scores for each dimension, resource gaps and sufficiency / critical / insufficiency status for each dimension, comprehensive suitability score and level, and warning labels and allocation suggestions when risks exist.

[0105] The synergistic effect of the entire process is reflected in the fact that, starting from the same requirement document, an auditable feature list is obtained through three-layer deconstruction, a differentiable structural representation is obtained through the feature interaction graph, a stable four-dimensional prediction is obtained through the end-to-end graph network, and a conclusion that can be directly incorporated into management decisions is obtained through the fitness formula. The entire process is recordable, reproducible, and incrementally updated.

[0106] like Figure 7 As shown, this application also provides an information project multi-granularity feature deconstruction and resource adaptability evaluation system. This system can be deployed on an application server or cloud platform in the form of software modules, and each module is connected to external project management databases, resource management systems and requirement document storage.

[0107] The system includes a text preprocessing module. This module acquires the requirements document for the project to be evaluated and performs text preprocessing to obtain standardized requirements text. Internally, it implements a pipeline corresponding to step S1, which removes irrelevant characters, segments sentences, segments words, and removes stop words. This pipeline can be encapsulated into a callable, standardized interface for use by other modules.

[0108] The system includes a multi-granularity feature deconstruction module. This module extracts project features from standardized requirement texts at three granularity levels: macro, meso, and micro, resulting in macro, meso, and micro feature sets. Internally, this module includes a text classification sub-model loader, a domain dictionary matching engine, a dependency parser, a functional module template matcher, a technology stack terminology matcher, a data architecture pattern recognizer, an interface type template matcher, a named entity recognition annotator, and a rule inference engine. The outputs of each sub-component are aggregated into a unified feature set structure, outputting a three-layer feature set with confidence levels.

[0109] The system includes a reference project feature library module. This module stores multi-granularity feature vectors of historically completed projects and their corresponding actual resource consumption data, supporting partitioned retrieval by project type and incremental updates. The underlying layer of this module can use a combination of a relational database and caching, with the partition index built on the project type field. It also provides co-occurrence frequency and conditional probability query interfaces for use by the feature interaction graph construction module.

[0110] The system includes a feature interaction graph construction module. This module constructs a multi-granularity feature interaction graph using multi-granularity features as nodes and inter-feature relationships as edges, including establishing intra-level and cross-level relationship edges and calculating edge weights. This module pulls statistical tables from the reference project feature library module, performs co-occurrence determination, mutual information calculation, word vector semantic similarity lookup, and conditional probability lookup, outputting a graph data structure in the form of an adjacency list for the inference module to consume.

[0111] The system includes a graph attention network inference module. This module loads pre-trained graph attention network model parameters, performs multi-head graph attention convolution and global attention pooling on multi-granularity feature interaction graphs, and outputs a fused feature embedding vector. This module encapsulates graph batching and forward propagation logic, exposing an inference interface from the input graph to the output embedding.

[0112] The system includes a multi-head resource prediction module. This module loads a pre-trained multi-head resource prediction network, embeds the fused features into a vector, passes it through a shared fully connected layer and four parallel resource prediction heads, and outputs the predicted demand quantities for personnel, time, equipment, and budget resources, respectively. This module receives the output vector from the inference module, executes a fully connected forward chain, and outputs a four-dimensional floating-point array.

[0113] The system includes a resource fit assessment module. This module calculates the fit between predicted resource demands and the available resource pool, obtaining fit scores and resource gaps for each dimension. It then generates a comprehensive fit rating based on preset weights, resulting in a resource fit assessment report. The module outputs a formatted structured report object, which can be serialized into JSON or rendered as PDF for project stakeholders to review.

[0114] The data flow between the above modules is as follows: the output of the text preprocessing module is connected to the multi-granularity feature deconstruction module; the output feature set of the multi-granularity feature deconstruction module is sent to the query interface of the reference project feature library module and the input of the feature interaction graph construction module; the output graph of the feature interaction graph construction module is sent to the graph attention network inference module; the output of the inference module is embedded and sent to the multi-head resource prediction module; the four-dimensional output of the multi-head resource prediction module is sent to the resource suitability evaluation module; and finally, the resource suitability evaluation module generates an evaluation report.

[0115] It should be noted that the above modules can be deployed centrally on a single server or split and deployed in a microservice manner. Text preprocessing and multi-granularity feature deconstruction can be placed in an asynchronous queue to process large batches of documents, while graph inference and prediction can be deployed on inference nodes with GPU acceleration to improve throughput.

[0116] This application achieves end-to-end automation from unstructured requirements documents to executable resource adaptation conclusions through the above methods and systems. It solves the problems of subjective inconsistency in traditional expert estimation methods and the problems of excessively coarse granularity and neglect of feature structure coupling in traditional analogy methods. This enables resource planning in information projects to have traceability, data-driven nature and closed-loop updatableness.

[0117] It should be noted that the terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates at least two. Unless otherwise stated, terms such as "front," "back," "left," "right," "lower," and / or "upper" are for illustrative purposes only and are not limited to a location or spatial orientation. Terms such as "comprising" or "including" indicate that the elements or objects preceding "comprising" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0118] The singular forms “a,” “the,” and “the” used in this application specification and appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0119] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects, characterized in that, include: Step S1: Obtain the requirements document of the project to be evaluated, and obtain standardized requirements text through text preprocessing; Step S2 involves multi-granularity feature deconstruction of the standardized requirement text: At the macro-granularity, project type, business domain, application scenario, and organizational scale are extracted through text classification and domain dictionary matching to form a macro-feature set; at the meso-granularity, functional modules, technology stack selection, data architecture type, and integration interface type are extracted through dependency parsing and template matching to form a meso-feature set; at the micro-granularity, interface definitions, data entities, functional atoms, and business rules are extracted through named entity recognition and rule reasoning to form a micro-feature set. Step S3: Construct a reference project feature library based on historical project data. Each record contains multi-granularity feature vectors of completed projects and their actual resource consumption data, and stores them in partitions according to project type. Step S4: Using all the features obtained in step S2 as nodes and the relationships between features as edges, construct a multi-granularity feature interaction graph; wherein, when the co-occurrence frequency of two features in the reference library within the same granularity exceeds a preset threshold, establish an intra-level association edge, and the weight is determined by the co-occurrence frequency and mutual information; when two features at different granularities have semantic relationships, establish a cross-level association edge, and the weight is determined by the semantic relationship strength and conditional probability. Step S5: Input the interaction graph into the pre-trained graph attention network, and after multiple layers of multi-head attention convolution and global attention pooling, obtain the fusion feature embedding vector representing the entire project. Step S6: Embed the fused features into the pre-trained multi-head resource prediction network, and output the demand forecasts for personnel, time, equipment and budget resources through a shared layer and four parallel prediction heads, respectively. Step S7: Compare the predicted resource demand with the available resource pool, calculate the fit score and resource gap for each dimension, and then weight them according to the preset weights to obtain the comprehensive fit score, thereby generating a resource fit score assessment report.

2. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S2, the macro-feature extraction process includes: inputting the standardized requirement text into a pre-trained text classification model to obtain classification results for project type and business domain; using a preset domain dictionary to perform keyword matching on the standardized requirement text to identify application scenario terms and organizational scale indicators; merging the text classification results with the dictionary matching results to remove duplicates, obtaining a macro-feature set, where each macro-feature contains feature terms and feature confidence, and the feature confidence is determined by the similarity between the prediction results of the text classification model and the dictionary matching results.

3. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S2, the extraction process of meso-level features includes: performing dependency parsing on the standardized requirement text to obtain the dependency relationship tree of each sentence; matching the dependency relationship tree using a predefined functional module template library to identify functional module features; identifying the technology stack selection features of the standardized requirement text using a predefined technology stack terminology library and regular expression patterns; identifying the data architecture type features of the standardized requirement text using a predefined data architecture pattern; identifying the integration interface type features of the standardized requirement text using a predefined interface type template; and merging the identification results of the above four types of features to obtain the meso-level feature set.

4. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S3, the construction process of the reference project feature library includes: for each completed project, extracting its multi-granularity feature vector and actual resource consumption vector; cleaning the feature vectors and resource consumption vectors of all completed projects, removing records with missing or obvious abnormalities in the resource consumption data; dividing the cleaned data into several partitions according to project type, with each partition arranged in order of project completion time; merging partitions with fewer projects than the preset requirement into other types of partitions; the reference project feature library supports incremental updates after adding new completed project data.

5. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S4, the edge weights are determined as follows: For intra-level related edges, the edge weights are determined by weighting the normalized standard mutual information and the normalized co-occurrence frequency. The normalized standard mutual information is calculated based on the probability of the feature appearing in the reference item feature library and the probability of co-occurrence. The normalized co-occurrence frequency is calculated based on the ratio of the number of items where the feature appears simultaneously to the total number of items. The two are combined using a preset balance coefficient. For cross-level related edges, the edge weights are determined by weighting the semantic similarity based on pre-trained word vectors and the conditional probability. The conditional probability refers to the probability that another feature also appears given that one feature appears. The two are combined using a preset balance coefficient.

6. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S5, the training process of the graph attention network model includes: using completed projects in the reference project feature library as training samples, constructing a multi-granularity feature interaction graph for each training sample as model input, and using the actual resource consumption vector of the project as supervision label; jointly training the graph attention network model and the multi-head resource prediction network end-to-end, using a weighted combination loss function as the training objective; using an optimizer for gradient backpropagation and employing a learning rate decay strategy; setting the number of training rounds and batch size; employing an early stopping strategy during training, stopping training when the validation set loss no longer decreases within several consecutive rounds, and saving the model parameters that perform best on the validation set; and dividing the dataset into a training set, a validation set, and a test set according to a preset ratio.

7. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S5, the node embeddings are updated through a multi-head attention mechanism in the first and second graph attention convolutional layers: for each attention head, the normalized attention coefficient of the neighboring nodes to the current node is calculated, and the features of the neighboring nodes are weighted and aggregated using this coefficient. After processing by the activation function, the outputs of all attention heads are concatenated to obtain the new node embedding of this layer. The normalized attention coefficient is obtained by concatenating the query vector and the key vector and then processing it with the activation function and normalization. Both the first and second graph attention convolutional layers have a preset number of attention heads and corresponding output dimensions, and the negative slope parameter of the activation function is set as needed. The global attention pooling layer performs a weighted summation of all node embeddings output by the second layer through learnable global attention weights, and concatenates the weighted summation result with the max pooling result to obtain the fused feature embedding vector.

8. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S6, the training loss function of the multi-head resource prediction network is a composite loss function, which is composed of a weighted combination of the mean squared error and the mean absolute percentage error of each resource dimension, and an L2 regularization term is added to the model parameters to prevent overfitting. The loss weights of each resource dimension are preset according to the importance of the resource, and the global error term and regularization coefficient are set empirically. The shared fully connected layer contains multiple fully connected networks, each followed by an activation function and a random dropout layer to prevent overfitting, and finally outputs a fixed-dimensional shared feature representation. The four resource prediction heads each independently contain multiple fully connected networks, each followed by an activation function, and the last layer has no activation function and directly outputs the single demand prediction of the corresponding resource.

9. The method for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects according to claim 1, characterized in that, In step S7, the generation of the resource suitability assessment report also includes: classifying the resource status of each dimension into three levels—sufficient resources, critical resources, and insufficient resources—based on the positive and negative direction and magnitude of the resource gap in each dimension, with each level determined according to a preset warning threshold; classifying the comprehensive suitability level into multiple preset levels based on the numerical range of the comprehensive suitability; and marking risk warning information and generating resource allocation suggestions in the assessment report for resource dimensions determined to be insufficient or projects with low comprehensive suitability levels.

10. A system for multi-granularity feature deconstruction and resource adaptability evaluation of information technology projects, characterized in that, include: The text preprocessing module is used to obtain the requirements document of the project to be evaluated and perform text preprocessing to obtain standardized requirements text. The multi-granularity feature deconstruction module is used to extract project features from standardized requirement texts at three granularity levels: macro, meso, and micro, to obtain macro feature sets, meso feature sets, and micro feature sets. The reference project feature library module is used to store multi-granularity feature vectors of historical completed projects and corresponding actual resource consumption data, and supports partitioned retrieval and incremental updates by project type. The feature interaction graph construction module is used to construct a multi-granularity feature interaction graph with multi-granularity features as nodes and inter-feature relationships as edges, including the establishment of intra-level and cross-level relationship edges and the calculation of edge weights. The graph attention network inference module is used to load pre-trained graph attention network models, perform multi-head graph attention convolution and global attention pooling on multi-granularity feature interaction graphs, and output fused feature embedding vectors. The multi-head resource prediction module is used to load a pre-trained multi-head resource prediction network, embedding the fused features into a vector through a shared fully connected layer and four parallel resource prediction heads, and outputting the demand predictions for personnel, time, equipment and budget resources respectively. The resource suitability assessment module is used to calculate the suitability between the predicted resource demand and the available resource pool, obtain the suitability scores and resource gaps for each dimension, and generate a comprehensive suitability level according to preset weights, thus forming a resource suitability assessment report.

Images