An academic graph data generation method and system for multi-model database query performance evaluation

Abstract

The application belongs to the technical field of database benchmark test, and discloses a data generation method and system for multi-model database query performance evaluation. A graph sampling algorithm based on probability propagation is used to extract a seed data set that retains topological features from a large-scale academic network to construct a probability distribution model. In the generation process, a unified logical entity object is constructed based on statistical characteristics, and four kinds of modal data, i.e. relational, document, graph structure and vector, are synchronously analyzed and mapped to ensure strong consistency of cross-model semantics. A specific domain statistical language model and deep semantic coding are introduced to realize semantic alignment of unstructured text and high-dimensional vectors. Through a closed-loop time sequence evolution mechanism, incremental features are fed back to the probability model in real time to drive the next time step generation. The application effectively solves the problems of existing generated data, such as loss of real data features, weak cross-model consistency and lack of time sequence causality, and provides a high-fidelity and scalable test benchmark for multi-model database performance evaluation.

Description

Technical Field

[0001] This invention relates to the field of database benchmarking and data management technology, and in particular to an academic graph data generation method and system for evaluating the query performance of multi-model databases. Background Technology

[0002] With the increasing complexity of big data applications, the data generated by various industries is experiencing explosive growth and diversification. Today, a single data model is insufficient to meet the diverse data storage and analysis needs; for example, relational databases suffer from performance limitations when processing complex graph queries. To address this issue, Multi-model Databases (MMDBs) significantly reduce system complexity by simultaneously supporting multiple data models within a single kernel or unified query engine, making them a key next-generation data management platform attracting significant attention from both academia and industry.

[0003] Benchmarking multi-model database systems relies on large-scale, high-quality, and structurally complex test data. Because real-world production data often involves privacy-sensitive information and is difficult to scale on demand, data generation technology has become an indispensable part of building controllable testing environments and driving the iteration of multi-model database technology. How to build a data generation tool that can simulate the complexity of the real world and meet the constraints of multi-model data associations has become an urgent need in the current database field.

[0004] Although several benchmarks for multi-model databases have emerged in recent years, aiming to evaluate a system's ability to manage data across models, they still have significant limitations in their core test data generation mechanisms and applicable environments, making it difficult to meet the needs of in-depth evaluation of complex business scenarios. For example, when facing large-scale testing requirements, M2Bench's data generator often uses simple data copying or basic interpolation algorithms for expansion, destroying the characteristics of real data; UniBench focuses on satisfying static feature constraints to generate data snapshots of a specific scale, lacking a time-series dynamic evolution mechanism based on historical statistical feedback, and cannot simulate the complex dynamic evolution processes in the real world. In addition, most existing multi-model database benchmarks use data from social e-commerce, healthcare, etc., as test datasets. When generating academic graphs, a representative heterogeneous information network data type, their data generation tools cannot adapt to the unique long-tail network topology and time-series evolution patterns, thus failing to generate high-quality, high-fidelity academic data that meets the stress testing requirements of multi-model databases.

[0005] Therefore, there is an urgent need for an academic graph data generation method based on real data statistics that is oriented towards multi-model database evaluation. Summary of the Invention

[0006] This invention proposes an academic graph data generation method and system for performance evaluation of multi-model database queries. It aims to address the problems of existing data generation systems for multi-model database benchmarking, which fail to preserve the distribution characteristics of academic graph data, lack realistic simulation of data temporal sequence, and suffer from semantic inconsistencies and high verification overhead when generating data in model order. This invention designs a sampling method based on elastic jumps in discrete statistical distributions, preserving the true data distribution characteristics while filling data gaps in the long-tail distribution features commonly found in academic graph data. It designs a bidirectional temporal expansion architecture with forward natural evolution and backward historical backfilling, covering the needs of multi-model databases throughout the entire data lifecycle. By constructing logical entity objects containing complete attributes and topological relationships, and synchronously parsing and mapping them to multiple model data, it ensures data consistency and zero verification overhead in multi-model testing.

[0007] The technical solution of this invention is as follows: A method for generating academic graph data for performance evaluation of multi-model database queries, comprising the following steps: S1. Obtain the full dataset of the original academic knowledge graph, extract seed datasets using a graph sampling method based on probability spread, construct a probability distribution rule base based on the seed dataset using an elastic jump method, and train an N-Gram language model library using the Kenlm tool; S2. Set configuration parameters, determine the generation mode, and calculate the scale of new entities to be generated at each time step under the current generation mode; S3. Execute the generation task of the newly added entity scale round at each time step, including entity sampling based on the probability distribution rule library, generating unstructured text based on the N-Gram language model library using a hybrid text generation strategy, performing vector embedding and topological relationship construction, and encapsulating all generation results into a unified logical entity object; S4. The logical entity object is parsed and synchronously mapped into four types of model data: relational, document, graph, and vector, according to the multi-model data schema specification, and stored in the corresponding data files; S5. Perform post-processing at fixed time steps, merge duplicate data, update the relevant attributes of the generated data, and provide feedback to update the probability distribution rule base until the preset termination condition is met.

[0008] Step S1 includes: S1.1 Obtain the full dataset of the original academic knowledge graph, select an initial seed article set, and sample articles from neighboring articles along the citation relationship with decreasing probability according to the number of hops to add them to the set, forming a seed dataset with a scaling factor of 1; the dataset is stored in CSV format, including basic metadata and topological associations of core entities; the core entities are divided into four categories: authors, articles, topics, and institutions, and are classified into four modes: relational data, document data, graph data, and vector data; S1.2 Read the seed dataset, statistically analyze the distribution characteristics of core entities, including time span, article type, language, topic hierarchy, number of authors, and number of citations, calculate the author activity index, and construct a probability distribution rule base based on the elastic jump method; the elastic jump method specifically includes: using the statistically obtained discrete distribution characteristics as anchor points, calculating the integer fault spacing between adjacent anchor points to quantify local sparsity, and assigning jump probabilities to sparse intervals with fault spacing based on the local sparsity; S1.3 Construct an author collaboration network graph and topic hierarchy tree, calculate the main research areas of each author based on their articles, and construct an index citation pool based on topic classification, which is then written into the probability distribution rule base. S1.4 For each topic or superior domain, extract its associated text corpus, train a domain-specific text generator based on the N-gram statistical language model, and establish a high-frequency vocabulary and grammatical constraints to form an N-gram language model library.

[0009] In step S2, the generation logic of the generation mode is based on a bidirectional temporal extension architecture and includes two modes: S2.1 Forward Natural Growth Mode: Simulates the academic data generation process over a future period. Based on the historical time series of the seed dataset and the set configuration parameters, it calculates the entity growth curve for future time steps and sets the probability of introducing new topics that dynamically decays over time. S2.2 Backward History Filling Mode: Simulates the retrospective completion of academic data during version iteration. Based on the set configuration parameters, the newly added entity quantity is proportionally allocated to each time step of the historical time series.

[0010] The task generated in step S3 is as follows: S3.1 Author Team Construction: Based on the author activity index, core authors are sampled, and an author team is constructed according to the author collaboration network graph; S3.2 Multi-topic hierarchical sampling: Based on the author's main research areas, weighted sampling of related topics is performed in the topic hierarchy tree; S3.3 Citation Relationship Construction: A hybrid sampling strategy is adopted, in which a part of the citations are selected from the index citation pool based on the priority connection mechanism, and another part of the citations are selected from articles with the same parent node in the topic hierarchy tree, so as to simulate the Matthew effect and random discovery characteristics of academic citations. S3.4, Article Attribute Generation: Based on the probability distribution rule base, sample and generate basic attributes of the article entity, call the N-Gram language model library, and use a hybrid text generation strategy to generate the title and summary of the entity in combination with the high-frequency vocabulary and grammatical constraints; S3.5 Vector Embedding Generation: Input the generated text content into a pre-trained deep learning semantic encoding model to generate high-dimensional feature vectors.

[0011] Step S4, which parses and synchronously maps data into four types of model data, specifically includes: S4.1 Extract the basic attributes and foreign key relationships of logical entity objects and map them to multiple flat table structure records in a relational database; S4.2 Parse hierarchical metadata, inverted index, and nested list structure from logical entity objects and encapsulate them into JSON format document-type data records; S4.3 Extract many-to-many relationships and attribute graph features from logical entity objects, construct vertex files and edge files, and map them to graph database structure records; S4.4 Extract the unique identifier and high-dimensional feature vector of the logical entity object, and map them to vector records supported by the vector database.

[0012] In step S5, the post-processing specifically includes: S5.1 At the end of each time step, aggregate and count the number of citations of newly added articles, newly generated cooperative relationship edges, and newly emerging topic entities within that time step, and update the related entity attributes in the global state; S5.2 At the end of the preset update interval time step, the incremental data is merged into the probability distribution rule base to realize dynamic evolution simulation with time causality.

[0013] An academic graph data generation system for multi-model database query performance evaluation, executing the academic graph data generation method for multi-model database query performance evaluation as claimed, includes: The pattern modeling module acquires the full dataset of the original academic knowledge graph, extracts seed datasets using a graph sampling method based on probability spread, constructs a probability distribution rule base based on the seed datasets using an elastic jump method, and trains an N-Gram language model library using the Kenlm tool; it also extracts seed datasets and statistically analyzes real-world patterns based on global states for subsequent generation. The incremental calculation module determines the generation mode and calculates the scale of new entities to be generated at each time step under the current generation mode, and calculates the generation plan based on the input parameters. The entity generation module executes the generation task of the newly added entity scale round at each time step, including entity sampling based on the probability distribution rule library, generating unstructured text based on the N-Gram language model library using a hybrid text generation strategy, performing vector embedding and topological relationship construction, and encapsulating all generation results into a unified logical entity object. The data mapping module parses and synchronously maps the logical entity object into four types of model data: relational, document, graph, and vector, according to the multi-model data schema specification, and stores them in the corresponding data files. The temporal evolution module performs post-processing at fixed time steps, merging duplicate data, updating the relevant attributes of the generated data, and updating the probability distribution rule base until the preset termination condition is reached; it also statistically analyzes and updates related information to achieve dynamic closed-loop feedback in the data generation process.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, this invention fills in the data gaps in the long-tail distribution characteristics commonly found in academic graph data while preserving the true distribution characteristics of the data. Existing data generation methods for multi-model database query performance testing are mostly designed for datasets such as social graphs and e-commerce. Therefore, existing data generation methods often employ prior-based distribution fitting or discrete normalization sampling applicable to these datasets when statistically analyzing true data characteristics and sampling. However, the data characteristics of academic graphs differ significantly from these datasets. Academic graph data often exhibits a discrete long-tail distribution and is frequently accompanied by irregular local peaks and data gaps. Therefore, this invention designs a sampling method based on the elastic jump of discrete statistical distributions, solving the problem that the irregular characteristics of academic graph data cannot be fitted based on prior distributions, and filling in the gaps in zero-probability statistical data. This provides benchmark test data for multi-model database query performance evaluation that combines microscopic absolute fidelity with long-tail generalization ability.

[0015] Secondly, this invention designs a bidirectional temporal expansion architecture that combines forward natural evolution and backward historical backfilling. Existing data generation methods for evaluating the performance of multi-model database queries, when simulating the temporal evolution of data, mostly employ unidirectional timeline extrapolation, which can only simulate newly added entities and relationships in a strictly sequential manner to accommodate the streaming append-only writing pattern of conventional datasets. However, the lifecycle of academic graphs not only includes strict temporal causal constraints but also, due to the specific maintenance and update patterns of academic databases, often involves retrospective completion of missing data from historical years and version iterations. Therefore, this invention considers both forward and backward time series to design a bidirectional data generation mode, reproducing the macroscopic evolutionary trend of the forward natural growth of academic graphs and simulating its reverse historical omission completion, covering the evaluation needs of multi-model databases throughout the entire data lifecycle.

[0016] Third, it constructs logical entity objects containing complete attributes and topological relationships. Existing data generation methods for evaluating the performance of multi-model database queries employ a single-model serial generation strategy when processing data from different models. For example, relational data is generated first, followed by graph edges or text attributes. Therefore, when outputting multi-model data, semantic fragmentation and data state inconsistency between models are prone to occur, and subsequent consistency checks between models require significant computational and I / O overhead. In this invention, at each generation step, a logical entity object containing complete attributes and topological relationships is first constructed in memory, and then synchronously converted into data from multiple models. This eliminates the data inconsistency problems that may result from step-by-step generation and completely saves on additional verification overhead.

[0017] Furthermore, thanks to the aforementioned sampling method based on the elastic jump of discrete statistical distribution and the bidirectional temporal extension architecture, this invention introduces a statistical feature feedback mechanism during the generation process. By abandoning the statistical feature method of fitting prior distributions, the probability distribution rule base can update its statistical features based on incremental data generated forward or backfilled backward, dynamically accumulating and correcting for distribution drift caused by special events (such as sudden academic hotspots) during the simulation's time progression. Compared to the data homogenization problem caused by the constraints of static statistical features on all subsequent time series generated data, this feedback mechanism enables the generated data to truly reflect the nonlinear influence of time series data on the evolution of subsequent data generation. Attached Figure Description

[0018] Figure 1 Flowchart of an academic graph data generation system for performance evaluation of multi-model database queries; Figure 2 A detailed flowchart for entity generation and data mapping; Figure 3The data expansion mode principle of this invention is as follows: (a) forward natural growth mode; (b) backward history backfill mode. Figure 4 A flowchart of an academic graph data generation method for performance evaluation of multi-model database queries. Detailed Implementation

[0019] To achieve the above objectives, the present invention adopts the following technical measures in terms of data feature statistical sampling, generation pattern framework construction, and multi-model data generation: Firstly, regarding the statistical sampling of data distribution characteristics, this invention designs a sampling method based on discrete statistical distribution elastic jump. In the feature statistics stage, the system performs discrete distribution statistics on the features of the real seed dataset of the academic graph, solidifying the real feature values ​​as anchor points to preserve irregular local peaks in the real data. The system calculates the numerical difference between adjacent anchor points, defined as the gap. In the data generation sampling stage, the system samples anchor points according to the original true probability. If a gap greater than 1 is detected on one side of an anchor point, elastic jump is performed on the gap feature values ​​within that gap, based on a weight that decreases inversely proportional to the distance between the feature value and the starting anchor point. Elastic jump is performed on the gaps between anchor points based on a weight that decreases inversely proportional to the distance, thus both sampling according to the real data characteristics and scientifically extrapolating and filling data gaps.

[0020] Secondly, regarding the framework for constructing the generation model, this invention designs a bidirectional temporal expansion architecture that includes forward natural evolution and backward historical backfilling. The forward natural evolution model, driven by behavioral simulation, strictly generates new academic entities and topological associations in ascending chronological order, aiming to simulate the macroscopic expansion of academic graph data in the future. The backward historical backfilling model, within the generated historical time interval, triggers reverse insertion and association reconstruction of data based on specific probabilities, thereby simulating historical omission completion and local topological encryption in real database maintenance. Furthermore, this invention introduces a statistical feature feedback mechanism into the aforementioned bidirectional temporal framework. After data generation at each time step is completed, the system will perform a re-normalization calculation and update of the probability distribution rule base according to the characteristics of the newly added logical entity objects, thereby restoring nonlinear patterns such as data distribution drift under long-period time series.

[0021] Thirdly, regarding multi-model data generation, this invention constructs logical entity objects containing complete attributes and topological relationships, and designs a single-synchronous multi-model mapping mechanism. Specifically, in the data generation phase, the system creates a logical entity object in memory for each generation round. This object, as an atomic data structure in memory, fully encapsulates all the generated data for that round: including basic scalar attributes (such as article publication year and article type), complex nested attributes (such as a hierarchical list of authors and their affiliated institutions), topological association sets (such as a list of reference IDs), unstructured text (such as paper abstracts), and embedded vectors. In the data instantiation phase, the system does not adopt a step-by-step serial generation strategy, but performs synchronous multi-path concurrent mapping. The system maps the logical entity object into data for multiple models according to preset mapping rules.

[0022] Through the above method, the present invention can generate test data that combines real data statistical features that fill data gaps, bidirectional time-series generation mode, and data consistency among multiple models. It makes up for the shortcomings of existing generation methods in generating academic graph data, such as micro-feature distortion, homogenization of static statistical features, and large cross-model consistency verification overhead. It provides reliable academic graph data for the performance evaluation of multi-model databases.

[0023] After comprehensively considering the above factors, this invention proposes an academic graph data generation method for evaluating the query performance of multi-model databases. Taking an Openalex open-source academic graph data snapshot as an example, the method includes the following steps: S1: Extract seed datasets from the original large-scale academic graph data and perform preprocessing for generation; S2: Initialize the generation tool; S3: Execute the generation task and encapsulate the generation results into a unified logical entity object; S4: According to the multi-model data schema specification, the logical entity is parsed and synchronously mapped into four types of model data: relational, document, graph structure and vector, and stored in the corresponding data files; S5: Statistically analyze the incremental data characteristics and feed them back to update the original probability distribution model. Apply the updated model to the generation of the next time step until the preset termination condition is met.

[0024] The specific process of step S1 includes: S1.1: Obtain a snapshot of the OpenAlex data. OpenAlex is an open academic literature dataset maintained by a non-profit organization. The dataset used for multi-model database benchmarking is stored in CSV format and contains basic metadata and topological relationships for four core entities: Authors, Works, Topics, and Institutions. Perform probability-based graph sampling, load the citation network of the original OpenAlex dataset, and build an in-memory adjacency list. The seed article is added to the sampling set, and a sampling probability is assigned to all connected articles in the citation network of the seed article. ,in The attenuation coefficient is... For the number of jumps, in probability Recursive spreading sampling is performed on citation edges. To prevent over-concentration of sampling, a maximum number of seed articles per sampling and a maximum spreading depth for a single seed article are set. Sampling stops when the total number of sampled articles reaches a preset threshold (e.g., the number of baseline articles set for SF=1). The sampled article nodes and their associated authors, topics, and institutional entities are extracted, and dangling edges are removed to form a closure seed dataset.

[0025] S1.2: Read the seed dataset and statistically analyze the distribution characteristics of the core entities, as shown in Table 1.

[0026] The formula for calculating the topic popularity score is as follows: This indicates the total number of related articles under this topic. This indicates the total number of citations for this topic. It is a local normalization operator, subscript This indicates that the reference range for normalization is limited to the subfield to which the topic belongs, rather than the entire dataset. These are preset weight parameters.

[0027] The formula for calculating author activity is: This indicates the total number of articles published by the author. This indicates the author's total number of citations. It is a global normalization operator. These are preset weight parameters.

[0028] Table 1 Statistical Characteristics

[0029] S1.3: Mining the relationship structure between entities and constructing a graph and inverted index pool to support complex sampling. The specific operations are as follows: 1. Construct an author collaboration network graph: Read the author collaboration edge file and construct an author collaboration adjacency table in memory, which is used to sample based on real social relationships when generating author teams; 2. Constructing Author Domain Inference and Mapping: By connecting the "author-article" and "article-topic" relationships, obtain the topics and sub-domains corresponding to all articles published by each author throughout their history. Calculate the publication frequency of each author in each sub-domain, take the mode as the author's main research domain, and construct an author-domain mapping dictionary; 3. Construct a hierarchical citation sampling pool: For each topic and subdomain, create an inverted index containing all articles under that topic and their citation counts. Topics are used for citation sampling, and subdomains are used for backoff when data is sparse; 4. Construct topic semantic mapping: Establish a mapping table from topics to their text keywords and titles, and construct a reverse lookup dictionary from topics to subdomains and from subdomains to domains to ensure that contextual semantic information can be accurately obtained when generating text and vectors.

[0030] S1.4: Article summaries from the OpenAlex dataset are grouped and trained according to their subject matter and parent domain. For each group's corpus, an N-gram language model library is trained using the KenLM tool. High-frequency vocabularies for each domain are compiled, and grammatical constraints are set to ensure the readability of the generated text.

[0031] The specific process of step S2 includes: This invention designs two data generation modes, determined by parameters in the generation tool configuration. The generation logic of each mode is based on a bidirectional temporal extension architecture and includes two modes: S2.1: Forward natural growth mode. The system first calculates the required total generation increment based on the scale factor (SF) in the configuration parameters and the initial total number of articles in the seed data. Then, the system uses the cutoff time of the seed data snapshot. Starting from a predetermined time step (e.g., years), move towards the future target time. An incremental extrapolation is performed. Specifically, for future years... Target generation quantity From the year Number of generation Combined with natural growth rate Expand and introduce noise. The perturbation calculation yields the following growth formula: A decay function was also set up to decrease the probability of introducing new topics each year as the simulation progresses, simulating the maturation process of a discipline. The probability of introducing new authors in each generation round was also set to simulate the influx of new authors into the academic community. The generation pattern is as follows: Figure 3 As shown in part (a).

[0032] S2.2: Backward Historical Data Filling Mode. The system calculates the total data increment required for the historical data filling stage based on the proportionality factor SF in the configuration parameters and the initial total number of articles in the seed data. It then calculates the data volume proportion of each time step in the seed dataset and distributes the total data increment to each specific time step according to this density proportion, determining the data filling task volume for each time step. Simultaneously, it keeps the number of authors and topics constant, aiming to simulate historical omission filling in real database maintenance. The generation mode is as follows: Figure 3 As shown in section (b).

[0033] The specific process of step S3 includes: The generation tool generates an article and related entities in each round and encapsulates them into a unified logical entity object.

[0034] S3.1: Invoke the author selector to construct the author team. Determine the number of authors for the article based on the truncated long-tail author number distribution in the article attribute distribution described in S1.2. Sample a "core author" based on the activity weight described in S1.2, and find the neighbor nodes of the core author from the cooperation network graph described in S1.3 to add them to the author team. If the number of selected authors is insufficient, supplement them by sampling from globally active authors.

[0035] S3.2: Perform multi-topic hierarchical sampling, determining the topic to which the article belongs by sampling the topic number distribution in the article attribute distribution. Based on the author team described in S3.1 and referring to the authors' main research areas described in S1.3, calculate the mode as the "anchor area". In the topic hierarchy tree, prioritize sampling the main topic from this anchor area, and sample sub-topics across areas with a certain probability to simulate interdisciplinary research.

[0036] S3.3: Call the citation selector to perform mixed sampling, which samples high-impact articles from the highly cited pool of the same topic by weighted sampling based on citation count, and randomly and uniformly samples from the pool of articles in the same field to form the reference list of the new article.

[0037] S3.4: Attribute Content Filling. For each entity, attributes are sampled and filled according to the content in Table 1. Other attributes are filled using a fixed generation logic. For example, `work.type` is sampled according to the article type statistics in Table 1, `work.language` is sampled according to the frequency distribution of the language field in Table 1, and `work.doi` is filled according to the format "https: / / doi.org / 10.{part1} / {part2}.{current_year}.{part3}". `part1`, `part2`, and `part3` are filled with randomly generated fixed-length numbers, and `current_year` is filled with the year assigned by the generator… For unstructured text in the abstract, this invention proposes a layered splicing generation strategy. First, the middle part of the abstract is generated. If the article has a subtopic, the system randomly selects a template from a pre-set cross-domain template library and fills in the subtopic name. If the article has cited keywords, the citation template library is called to fill in the keywords. The remaining space after subtracting the middle layer length from the total length of the target abstract is calculated and allocated proportionally to the beginning and end. The N-gram statistical language model corresponding to the main topic is called to generate the beginning and end text respectively. During the word-by-word generation process of the N-gram model, a hard-constraint rule library is introduced for real-time filtering to ensure that the generated text conforms to academic writing conventions in terms of grammatical structure. The first sentence of the complete abstract is extracted, and the ending punctuation is removed to serve as the article's title. Additionally, a non-structured text attribute generation mode with random character padding is provided, determined by parameters in the generation tool configuration.

[0038] S3.5: Generate vectors for the word embeddings of the generated unstructured text. Based on the defined data pattern, the content in the vector model corresponds to an article obtained by concatenating the title and abstract and then embedding them using a language model. The system initializes a pre-trained deep learning semantic encoding model (in this embodiment, the lightweight model all-MiniLM-L6-v2 under the SentenceTransformer framework is used). The model is loaded using CPU inference mode, and a singleton pattern ensures that each worker process loads the model instance only once. The article title and article abstract described in S3.4 are concatenated to form a complete semantic context input. The concatenated text is input into the encoding model, forward inference is performed, and a 384-dimensional fixed-dimensional vector is output. This vector array is then converted to string format for subsequent writing to a CSV file and parsing and importing into a vector database.

[0039] The specific process of step S4 includes: The system, based on the logical entity object which contains all the information of an academic entity, uses a specific serializer to decompose the same logical data and synchronously map it to four different physical storage formats.

[0040] S4.1: Relational data mapping generates structured data conforming to the third normal form. The system extracts scalar attributes from logical entity objects, filtering out nested lists and complex objects. Within the defined OpenALEX data schema, the relational data generator maps the contents of two files: works.csv and authors.csv. Each element in the works.csv relational table contains 12 attributes: id, doi, title, display_name, publication_year, publication_date, type, cited_by_count, is_retracted, is_paratext, cited_by_api_url, and language. Each element in the authors.csv relational table contains 8 attributes: id, display_name, works_count, cited_by_count, last_known_institution, works_api_url, updated_date, and institution_id. The system assembles these fields into records according to a fixed column order and writes them to the corresponding files.

[0041] S4.2: Document-based data mapping, generating semi-structured, hierarchically nested data. The system preserves the related data in logical entities as a nested structure, constructs author information into an `authorships` array, where each element contains attributes such as `author_id`, `author_position`, and `institution`. It constructs a `topics` array for topic information, where each element contains attributes such as `id`, `display_name`, and `score`. These, along with attributes such as `language`, `abstract`, and `volume`, constitute the JSON fields of the `doc` document. Combined with the `id`, this forms a document element for an article and is written to `works_doc.csv`. Simultaneously, the system constructs the `orcid`, `display_name_alternatives`, and other fields into the JSON fields of the `doc` document, combining them with the author's `id` to form a document element for an author and writing it to `authors_doc.csv`.

[0042] S4.3: Graph data mapping, extracting topological connections between entities. The graph data in the specified data schema is stored in CSV files containing vertex and edge tables. Each vertex file contains a corresponding ID and a properties field, stored in JSON format. Edge files contain startid, endid, and properties in JSON format, as detailed in Table 2.

[0043] Table 2. Data storage format.

[0044] S4.4: Vector data mapping, directly obtain the feature vector array generated in S3.5, and write the entity ID and vector string into works_vec.csv.

[0045] The specific process of step S5 includes: S5.1: At the end of each time step, the system scans the newly added data in the temporary buffer, reads the newly added citation relationship edge files for the current year, aggregates and counts the cited articles, and obtains the number of new citations for each historical article in the current year; it reads the newly added article author relationship edge files and article topic edge files for this time step, and counts the number of new publications for each author and the number of new related articles for each topic; based on the new publication information of related authors, it synchronously calculates the scale of new output for each research institution in the current year. The incremental values ​​obtained from the above aggregation are added to the cited_by_count and works_count attributes of the corresponding entity. For newly emerging topic entities, they are appended to the global topic list and their popularity value is initialized. After the update is completed, the latest status table is persisted, completing the data solidification for the current time step.

[0046] S5.2: When the preset asynchronous time interval is reached, the incremental data is merged into the global statistical state. The system calls the pattern modeling module and, based on the full dataset updated in S5.1, re-executes steps S1.2 and S1.3. Before proceeding to the next time step, the generator reloads the recalculated probability distribution parameters and index dictionary to update the probability distribution rule base. Through this mechanism, the generated dataset can reflect the nonlinear distribution drift over long periods. After all time steps are generated, the generated data files are stored in the output directory.

Claims

1. A method for generating academic graph data for performance evaluation of multi-model database queries, characterized in that, Includes the following steps: S1. Obtain the full dataset of the original academic knowledge graph, extract seed datasets using a graph sampling method based on probability spread, construct a probability distribution rule base based on the seed dataset using an elastic jump method, and train an N-Gram language model library using the Kenlm tool; S2. Set configuration parameters, determine the generation mode, and calculate the scale of new entities to be generated at each time step under the current generation mode; S3. Execute the generation task of the newly added entity scale round at each time step, including entity sampling based on the probability distribution rule library, generating unstructured text based on the N-Gram language model library using a hybrid text generation strategy, performing vector embedding and topological relationship construction, and encapsulating all generation results into a unified logical entity object; S4. The logical entity object is parsed and synchronously mapped into four types of model data: relational, document, graph, and vector, according to the multi-model data schema specification, and stored in the corresponding data files; S5. Perform post-processing at fixed time steps, merge duplicate data, update the relevant attributes of the generated data, and provide feedback to update the probability distribution rule base until the preset termination condition is met.

2. The academic graph data generation method for performance evaluation of multi-model database queries according to claim 1, characterized in that, Step S1 includes: S1.1 Obtain the full dataset of the original academic knowledge graph, select an initial seed article set, and sample articles from neighboring articles along the citation relationship with decreasing probability according to the number of hops to add them to the set, forming a seed dataset with a scaling factor of 1; the dataset is stored in CSV format, including basic metadata and topological associations of core entities; the core entities are divided into four categories: authors, articles, topics, and institutions, and are classified into four modes: relational data, document data, graph data, and vector data; S1.2 Read the seed dataset, statistically analyze the distribution characteristics of core entities, including time span, article type, language, topic hierarchy, number of authors, and number of citations, calculate the author activity index, and construct a probability distribution rule base based on the elastic jump method; the elastic jump method specifically includes: using the statistically obtained discrete distribution characteristics as anchor points, calculating the integer fault spacing between adjacent anchor points to quantify local sparsity, and assigning jump probabilities to sparse intervals with fault spacing based on the local sparsity; S1.3 Construct an author collaboration network graph and topic hierarchy tree, calculate the main research areas of each author based on their articles, and construct an index citation pool based on topic classification, which is then written into the probability distribution rule base. S1.4 For each topic or superior domain, extract its associated text corpus, train a domain-specific text generator based on the N-gram statistical language model, and establish a high-frequency vocabulary and grammatical constraints to form an N-gram language model library.

3. The academic graph data generation method for performance evaluation of multi-model database queries according to claim 2, characterized in that, In step S2, the generation logic of the generation mode is based on a bidirectional temporal extension architecture and includes two modes: S2.1 Forward Natural Growth Mode: Simulates the academic data generation process over a future period. Based on the historical time series of the seed dataset and the set configuration parameters, it calculates the entity growth curve for future time steps and sets the probability of introducing new topics that dynamically decays over time. S2.2 Backward History Filling Mode: Simulates the retrospective completion of academic data during version iteration. Based on the set configuration parameters, the newly added entity quantity is proportionally allocated to each time step of the historical time series.

4. The academic graph data generation method for performance evaluation of multi-model database queries according to claim 3, characterized in that, The task generated in step S3 is as follows: S3.1 Author Team Construction: Based on the author activity index, core authors are sampled, and an author team is constructed according to the author collaboration network graph; S3.2 Multi-topic hierarchical sampling: Based on the author's main research areas, weighted sampling of related topics is performed in the topic hierarchy tree; S3.3 Citation Relationship Construction: A hybrid sampling strategy is adopted, in which a part of the citations are selected from the index citation pool based on the priority connection mechanism, and another part of the citations are selected from articles with the same parent node in the topic hierarchy tree, so as to simulate the Matthew effect and random discovery characteristics of academic citations. S3.4, Article Attribute Generation: Based on the probability distribution rule base, sample and generate basic attributes of the article entity, call the N-Gram language model library, and use a hybrid text generation strategy to generate the title and summary of the entity in combination with the high-frequency vocabulary and grammatical constraints; S3.5 Vector Embedding Generation: Input the generated text content into a pre-trained deep learning semantic encoding model to generate high-dimensional feature vectors.

5. The academic graph data generation method for performance evaluation of multi-model database queries according to claim 4, characterized in that, Step S4, which parses and synchronously maps data into four types of model data, specifically includes: S4.1 Extract the basic attributes and foreign key relationships of logical entity objects and map them to multiple flat table structure records in a relational database; S4.2 Parse hierarchical metadata, inverted index, and nested list structure from logical entity objects and encapsulate them into JSON format document-type data records; S4.3 Extract many-to-many relationships and attribute graph features from logical entity objects, construct vertex files and edge files, and map them to graph database structure records; S4.4 Extract the unique identifier and high-dimensional feature vector of the logical entity object, and map them to vector records supported by the vector database.

6. The academic graph data generation method for performance evaluation of multi-model database queries according to claim 1, characterized in that, In step S5, the post-processing specifically includes: S5.1 At the end of each time step, aggregate and count the number of citations of newly added articles, newly generated cooperative relationship edges, and newly emerging topic entities within that time step, and update the related entity attributes in the global state; S5.2 At the end of the preset update interval time step, the incremental data is merged into the probability distribution rule base to realize dynamic evolution simulation with time causality.

7. An academic graph data generation system for evaluating the query performance of multi-model databases, characterized in that, The academic graph data generation method for performance evaluation of multi-model database queries as described in any one of claims 1-6 includes: The pattern modeling module acquires the full dataset of the original academic knowledge graph, extracts seed datasets using a graph sampling method based on probability spread, constructs a probability distribution rule base based on the seed datasets using an elastic jump method, and trains an N-Gram language model library using the Kenlm tool; it also extracts seed datasets and statistically analyzes real-world patterns based on global states for subsequent generation. The incremental calculation module determines the generation mode and calculates the scale of new entities to be generated at each time step under the current generation mode, and calculates the generation plan based on the input parameters. The entity generation module executes the generation task of the newly added entity scale round at each time step, including entity sampling based on the probability distribution rule library, generating unstructured text based on the N-Gram language model library using a hybrid text generation strategy, performing vector embedding and topological relationship construction, and encapsulating all generation results into a unified logical entity object. The data mapping module parses and synchronously maps the logical entity object into four types of model data: relational, document, graph, and vector, according to the multi-model data schema specification, and stores them in the corresponding data files. The temporal evolution module performs post-processing at fixed time steps, merging duplicate data, updating the relevant attributes of the generated data, and updating the probability distribution rule base until the preset termination condition is reached; it also statistically analyzes and updates related information to achieve dynamic closed-loop feedback in the data generation process.

Images