An academic journal intelligent recommendation method and system based on multi-source heterogeneous data fusion

Abstract

The application discloses a kind of based on multi-source heterogeneous data fusion academic journal intelligent recommendation method and system, belong to academic information processing technical field.The method includes: receiving user paper information and preference setting;Call big language model API and extract structured paper features;Parallelly collect multiple academic open data source data and unify standardization;Through theme accurate matching and subject fuzzy matching double-channel recall candidate journal;Realize the reverse recommendation from paper to journal based on co-citation network;Five-dimensional weighted score of candidate journal is carried out theme matching degree, subject matching degree, influence matching degree, peer review cost matching degree and paper publishing evidence degree;Risk classification, OA type processing and early warning journal marking are carried out;Generate recommendation reason based on evidence citation, return Top-N recommendation result.The application solves the technical problems of single data source, limited matching dimension, lack of reverse recommendation path and risk assessment in the prior art.

Description

Technical Field

[0001] This invention relates to the field of academic information processing technology, and in particular to an intelligent recommendation method and system for academic journals based on the fusion of multi-source heterogeneous data. Background Technology

[0002] With the continuous growth of academic publishing, the number of academic journals worldwide has exceeded tens of thousands. After completing their papers, researchers need to select suitable journals for submission from among numerous options. The journal selection process involves considerations from multiple dimensions, including the relevance of the paper's topic, impact factor, review period, open access fees, and acceptance rate. Traditional manual journal selection methods are inefficient, often requiring researchers to spend a significant amount of time researching journal information.

[0003] Currently, there are various journal recommendation tools on the market, mainly including: text matching tools based on a single publisher's database (such as Elsevier Journal Finder, Springer Journal Suggester, JANE, etc.), query and filtering platforms based on journal databases (such as LetPub, ResearchPass, etc.), and academic recommendation systems based on collaborative filtering or knowledge graphs.

[0004] Existing technologies share the following common problems: Single data source: Existing tools are all based on a single data source (paper text, historical statistics, or a single API), none of which involve the integration of multiple heterogeneous academic data sources, failing to comprehensively cover major global academic journals; Limited matching dimensions: Existing technologies mainly rely on text similarity or feature matching from machine learning models, lacking multi-dimensional matching signals such as topic graph matching, co-citation network analysis, and evidence from published papers; Lack of reverse recommendation paths: Existing technologies are all forward matching from "paper to journal," lacking a reverse recall path of "paper → similar paper → journal," and cannot utilize the actual publication destinations of similar published papers as a basis for recommendations; Lack of risk assessment: Existing technologies lack a systematic integration of multi-dimensional risk assessment, such as grading the proportion of publications by Chinese authors, standardizing OA types, and marking early warning journals. In the academic publishing field, an excessively high proportion of authors from a single country is widely considered by the academic community to be an important indicator of journal quality risk. According to the National Science Foundation (NSF) report "Science & Engineering..." According to the Indicators report, China accounts for approximately 20-25% of the global output of academic papers. When the proportion of Chinese authors publishing in a journal deviates significantly from this global average, the journal may face risks such as a monolithic source of manuscripts and a decline in the quality of peer review. The journal warning list released by the Documentation and Information Center of the Chinese Academy of Sciences and New Academic Research also uses excessively high concentration of author sources as one of the important reference indicators for warning judgment. However, current technologies have not systematically integrated this indicator into the journal recommendation process. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent recommendation method and system for academic journals based on the fusion of multi-source heterogeneous data, so as to solve the technical problems of single data source, limited matching dimensions, lack of reverse recommendation path and lack of risk assessment in the existing technology.

[0006] To achieve the above objectives, this invention provides an intelligent recommendation method for academic journals based on multi-source heterogeneous data fusion, comprising the following steps: S1: Receive the user's paper information and preference settings, where the paper information includes the paper title, abstract and keywords, and the preference settings include the expected impact factor range, expected review period and expected databases; S2: Call the large language model API to extract features from the paper text and obtain structured paper features, which include keyword list, main subject classification, sub-subject direction and research type; S3: Based on the structured paper features obtained in step S2, perform multi-source heterogeneous data collection in parallel, call multiple academic open data source interfaces to obtain journal metadata, paper statistics and semantic search results respectively, and adopt a parallel asynchronous mode so that the failure of any data source call does not block the call of other data sources. S4: Convert heterogeneous data returned from various data sources into an internal standard schema through a unified data standardization module to achieve unified field mapping and data structure; S5: Parallel execution of dual-channel candidate journal recall, including: S5a: The topic precise matching channel, based on the subject topic map, calculates the overlap between the paper keywords and the journal topic, and recalls candidate journals by sorting the number of overlaps; S5b: Subject-based fuzzy matching channel, which performs fuzzy search based on multiple sets of mainstream subject classification fields to recall candidate journals; S6: Select seed papers from semantic search results, expand the co-citation network using the literature association field, obtain relevant papers and count their publication journal frequency, add high-frequency journals to the candidate set as a co-citation network enhancement channel, and realize the reverse recommendation path from paper to journal. S7: Merge and deduplicate the recall results from the three channels in steps S5 and S6, and sort them by topic matching degree to obtain a set of candidate journals; S8: Perform a five-dimensional weighted scoring on the candidate journals and calculate the comprehensive matching score (Score). The five-dimensional weighted scoring formula is as follows: Score = w1 x S_topic + w2 x S_subject + w3 x S_impact + w4 x S_cost +w5 x S_evidence Where S_topic is the topic matching degree, S_subject is the subject matching degree, S_impact is the impact matching degree, S_cost is the peer review fee matching degree, S_evidence is the paper publication evidence degree, w1 to w5 are the corresponding weight coefficients, and S_evidence is calculated based on the number of publications and citations of similar papers obtained from the co-citation network enhancement channel in step S6 in each candidate journal. S9: Conduct multi-dimensional risk assessments on the recommendation results, including risk grading based on the proportion of articles published by Chinese authors, standardization of OA types, and marking of warning journals, and outputting structured risk warnings; S10: Based on scores across various dimensions and evidence from data sources, generate fact-based recommendation reasons for each recommended journal and return Top-N recommendation results with matching scores, recommendation reasons, and risk warnings.

[0007] Furthermore, the multi-source heterogeneous data acquisition described in step S3 specifically includes: Retrieve journal topic graphs, h-index, i10-index, OA status, APC fees, and publication statistics of Chinese authors from the OpenAlex API; Retrieve journal publication volume, citation statistics, annual trends, and paper type distribution data from the Crossref API; Obtain semantic search results, TLDR summaries, high-impact citation tags, and similar paper recommendation data from the Semantic Scholar API.

[0008] Furthermore, the weight coefficients mentioned in step S8 satisfy the following priority relationship: w1>w2 ≥ w5>w3>w4, where w1 is the topic matching weight, which takes the maximum value in the five-dimensional score, w2 is the subject matching weight, w5 is the paper publication evidence weight, w3 is the influence matching weight, and w4 is the peer review fee matching weight, which takes the minimum value.

[0009] The priority relationship of the weighting coefficients is determined based on the degree of influence of each decision dimension on the submission decision in the journal recommendation scenario: The degree of topic relevance has the greatest impact on submission decisions, because matching the research direction of the paper with the scope of journal coverage is the primary condition for journal selection. Subject matching degree and paper publication evidence degree are the next best, providing matching basis from the two perspectives of subject classification and the actual publication destination of published papers, respectively; Impact matching degree is the next most important, reflecting the match between the journal's academic influence and users' expectations; Peer review fees have the least impact, as fees are usually not the decisive factor in journal selection.

[0010] Furthermore, the risk classification based on the proportion of articles published by Chinese authors mentioned in step S9 specifically refers to: We analyzed data on articles published by Chinese authors over the past five years and calculated the percentage of articles published by Chinese authors. A baseline is determined based on the proportion of natural publications by Chinese authors in global academic output. When the proportion is lower than the baseline, it is marked as a safe level. When the percentage is higher than the baseline but does not exceed the preset threshold for friendliness to Chinese people, it is marked as a friendliness level to Chinese people; When the percentage exceeds the preset threshold for friendliness towards Chinese people, it is marked as a high-risk level; When the proportion exceeds the preset high-risk threshold, it is marked as high-risk. When the proportion exceeds the preset extremely high risk threshold, it is marked as an extremely high risk level; The baseline mentioned above is determined based on the historical publication percentage of Chinese authors in global academic paper databases.

[0011] Furthermore, the baseline is taken as the statistical average of the proportion of publications by Chinese authors in global academic paper databases over the past 5 years, approximately 20-25%; the "Chinese-friendly" threshold is 1.5 to 2 times the baseline; the "high-risk" threshold is 2.5 to 3 times the baseline; and the "extremely high-risk" threshold is 3 to 3.5 times the baseline. This classification method references the assessment approach regarding author source concentration in the journal warning list published by the Documentation and Information Center of the Chinese Academy of Sciences and New Academic Journals, as well as the DOAJ's evaluation criteria for journal internationalization.

[0012] Furthermore, the OA type standardization process described in step S9 specifically includes: forcibly labeling Nature, Science, and Cell journals as hybrid type; performing fallback mapping on non-standard OA types, where green is mapped to closed and bronze is mapped to hybrid; inferring the DOAJ inclusion status based on the OA type, where gold and diamond types are marked as included, and closed and hybrid types are marked as not included.

[0013] Furthermore, the selection of seed papers from semantic search results in step S6 specifically involves selecting the top N papers with the highest relevance in the semantic search results as seed papers, where N ranges from 2 to 5; the expansion of the co-citation network specifically involves obtaining the relatedWorks list for each seed paper, batch obtaining the metadata of related papers, counting the number of related papers included in each journal and the cumulative number of citations, and adding journals with more than a preset threshold to the candidate set.

[0014] Furthermore, the merging and deduplication described in step S7 specifically involves: using the journal's unique identifier as the deduplication key to deduplicate the recall results from the three channels; for journals that are recalled by multiple channels simultaneously, retaining their highest topic matching degree as the sorting criterion; the upper limit of the candidate set is 100 journals.

[0015] This invention also provides an intelligent recommendation system for academic journals based on multi-source heterogeneous data fusion, comprising: The user input layer is used to receive users' paper information and preference settings; The feature extraction layer, connected to the user input layer, extracts features from the paper text based on the large language model API to obtain structured paper features; The multi-source data fusion layer includes several academic open data source modules and a unified data standardization module. The unified data standardization module is used to convert the heterogeneous data returned by each data source module into an internal standard schema. The three-channel recall layer includes a topic precise matching channel, a subject fuzzy matching channel, and a co-citation network enhancement channel, as well as a candidate set and deduplication module for merging and deduplicating the recall results of the three channels. The co-citation network enhancement channel selects seed papers from semantic search results, expands the co-citation network using document association fields, and counts the publication frequency of related papers in journals to realize a reverse recommendation path from papers to journals. The scoring and ranking layer includes a five-dimensional weighted scoring module, a risk assessment module, and a recommendation reason generation module. The five-dimensional weighted scoring module calculates a comprehensive matching score based on five dimensions: topic matching degree, discipline matching degree, influence matching degree, peer review fee matching degree, and paper publication evidence degree. The paper publication evidence degree is calculated based on similar paper publication data obtained through the co-citation network enhancement channel. The recommended output layer is used to output Top-N recommendation results with matching scores, recommendation reasons, and risk warnings.

[0016] Furthermore, the risk assessment module includes: a risk classification submodule for the proportion of publications by Chinese authors, used to calculate the proportion of publications by Chinese authors based on OpenAlex's statistics of publications by Chinese authors over the past 5 years, and to classify the risk into five levels according to a preset threshold; an OA type standardization submodule, used to standardize the OA types returned by OpenAlex, including top journal coverage rules, non-standard type fallback mapping, and DOAJ inclusion status inference; and a warning journal marking submodule, used to check whether a journal is in the warning journal list and add a mark. Attached Figure Description

[0017] Figure 1 This is a system architecture diagram provided for an embodiment of the present invention. The system consists of a six-layer architecture: The user input layer is located at the top of the system and includes four input modules: paper title, paper abstract, keywords, and user preferences. It is responsible for receiving the user's paper information and preference settings. The feature extraction layer is located below the user input layer. It includes an AI feature extraction module, as well as three sub-modules: keyword extraction, subject classification, and research type identification. It is responsible for converting unstructured paper text into structured feature vectors. The multi-source data fusion layer is located to the right of the feature extraction layer. It includes several academic open data source modules (in the preferred embodiment, it includes three parallel data source modules: OpenAlex, Crossref, and Semantic Scholar), as well as a unified data standardization module. The data source modules are called in parallel and asynchronously, and the unified data standardization module converts heterogeneous data into an internal standard schema. The three-channel recall layer is located in the middle of the system. It includes three parallel recall channels: topic exact matching channel, subject fuzzy matching channel, and co-citation network enhancement channel, as well as a candidate set deduplication module. The recall results of the three channels are merged and deduplicated to output a candidate journal set. The co-citation network enhancement channel selects seed papers from semantic search results, expands the co-citation network using document association fields, and counts the publication frequency of related papers in journals to realize a reverse recommendation path from paper to journal. The scoring and ranking layer is located below the three-channel recall layer and includes a five-dimensional weighted scoring module, a risk assessment module, and a recommendation reason generation module. The five-dimensional weighted scoring module calculates a comprehensive matching score based on five dimensions: topic matching degree, subject matching degree, impact matching degree, peer review fee matching degree, and paper publication evidence degree. The weights follow the priority relationship: topic matching degree > subject matching degree ≥ paper publication evidence degree > impact matching degree > peer review fee matching degree. The risk assessment module includes a risk classification sub-module based on the proportion of publications by Chinese authors (using the global proportion of publications by Chinese authors as a baseline for five-level classification), an OA type standardization sub-module, and a warning journal marking sub-module. The recommendation output layer is located at the bottom layer of the system, and outputs Top-N recommendation results, matching scores, recommendation reasons, and risk warnings.

[0018] The data flow is as follows: the paper information from the user input layer flows into the feature extraction layer, and the feature extraction results flow into the three-channel recall layer; the standardized data from the multi-source data fusion layer flows into the three-channel recall layer and the scoring and ranking layer; the candidate set from the three-channel recall layer flows into the scoring and ranking layer; and the final result from the scoring and ranking layer flows into the recommendation output layer.

[0019] Figure 2 The recommended flowchart provided for the embodiments of the present invention corresponds to steps S1 to S10 of claim 1, and the flow is as follows from top to bottom: Step S1: The user submits paper information, including the paper title, abstract, keywords, and preferred conditions such as the expected impact factor range and review period. Step S2, AI feature extraction: The system calls the large language model API to extract structured features from the paper text, including keyword list, main discipline classification, sub-discipline direction, and research type. Step S3: Parallel multi-source heterogeneous data acquisition is performed by calling multiple academic open data source interfaces to obtain journal metadata, paper statistics and semantic search results respectively; Step S4: Convert the heterogeneous data returned by each data source into an internal standard schema through the unified data standardization module; Step S5 involves parallel execution of dual-channel candidate journal recall, including S5a topic precise matching channel (based on subject topic map) and S5b subject fuzzy matching channel (based on multiple sets of mainstream subject classification fields). Step S6: Enhance the co-citation network channel by selecting seed papers from the semantic search results, expanding the co-citation network using the literature association field, and counting the publication frequency of related papers in journals to achieve a reverse recommendation path from papers to journals. Step S7: Merge and deduplicate the recall results from the three channels, sort them by topic matching degree, and obtain a set of candidate journals. Step S8: Perform a five-dimensional weighted scoring on the candidate journals and calculate the comprehensive matching score. The weights satisfy the priority relationship w1>w2≥w5>w3>w4. Step S9 involves conducting a multi-dimensional risk assessment of the recommendation results, including risk grading based on the proportion of publications by Chinese authors (with the global natural publication rate of approximately 25% as a baseline), standardized processing of OA types, and marking of warning journals. Step S10 generates fact-based recommendation reasons for each recommended journal and returns Top-N recommendation results with matching scores, recommendation reasons, and risk warnings. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Example 1: This example provides an intelligent recommendation method for academic journals based on multi-source heterogeneous data fusion, which is applied to academic journal recommendation scenarios.

[0022] Step S1: User Input Stage. Users input their paper title, abstract (optional), keywords (optional), desired impact factor range (e.g., 3-8), review period (e.g., no more than 3 months), and preferred databases (e.g., SCI, EI, etc.) through the front-end interface. The system receives and verifies the validity of the input parameters.

[0023] Step S2: AI Feature Extraction Stage. The system calls a large language model API (such as the DeepSeek API), taking the paper title and abstract as input. Structured prompts guide the model to extract the following features: a list of core academic keywords (5-10), the main subject category (e.g., computer science, medicine), the sub-subject direction (e.g., deep learning, medical imaging), and the research type (clinical research, experimental research, theoretical research, review, etc.). This step transforms the unstructured paper text into a structured feature vector that can be used for subsequent matching.

[0024] Step S3: Multi-source heterogeneous data acquisition stage. Based on the structured paper features obtained in Step S2, the system calls three academic data source APIs in parallel: OpenAlex API: to obtain metadata such as journal topic graphs, h-index, i10-index, OA status, APC fees, and publication statistics of Chinese authors. OpenAlex provides metadata and citation relationships for more than 250 million academic papers; Crossref API: to obtain statistical data such as journal publication volume, citation statistics, annual trends, and paper type distribution. Crossref covers more than 150 million DOI registration records; Semantic Scholar API: to obtain semantic search results, TLDR abstracts, influential citation counts, and similar paper recommendations. Semantic Scholar provides semantic-level paper similarity calculation based on the SPECTER2 model.

[0025] The three data sources are invoked in a parallel asynchronous mode. If any data source fails, it will not block the invocation of other data sources, thus ensuring the robustness of the system.

[0026] Step S4: Unified Data Standardization Phase. The three data sources return data with different structures. The system defines a unified internal data schema (JournalEnhancedSchema) to map fields from each data source to a unified structure. Specific mapping rules include: OpenAlex's h_index and i10_index are mapped to the unified hIndex and i10Index; Crossref's total_works and works_3years are mapped to the unified totalWorks and works3Years; Semantic Scholar's citationCount and influentialCitationCount are mapped to the unified citationCount and influentialCitationCount. Through the unified data standardization module, the structure of heterogeneous data is unified for use by downstream modules.

[0027] Step S5: Dual-channel candidate journal recall phase. The system executes two recall channels in parallel: S5a: Exact topic matching channel. The system obtains a list of topic IDs corresponding to the paper's keywords from the OpenAlex Topics API, and then searches for journals containing these topics in the local JournalOpenalexMeta table, sorting them by topic overlap for recall. This channel is based on the OpenAlex Topics map and can accurately match the research topic of the paper; S5b: Fuzzy subject matching channel. Based on the paper's primary and secondary subjects, the system performs fuzzy searches in multiple fields such as primary_subject, secondary_subjects, jcr_subject, xr_category, and cas_category in the Journal table to recall candidate journals. This channel is based on a multi-dimensional subject classification system (JCR subjects, Chinese Academy of Sciences classification, emerging journal category) and can capture matching signals at the subject level.

[0028] Step S6: Co-citation Network Enhancement Stage. The system selects the top 3 highly relevant papers from the Semantic Scholar semantic search results as seed papers and obtains a list of related papers associated with each seed paper through the relatedWorks field of the OpenAlex API. The system batch retrieves the details of these related papers, counts the frequency of their publication journals, and adds high-frequency journals to the candidate set. This channel realizes a reverse recommendation path from "paper to similar paper to journal," using the actual publication destination of published papers as the recommendation basis. This is the core innovation of this invention, which distinguishes it from existing technologies: existing technologies are all forward matching from "paper to journal," while this invention realizes a reverse recommendation path through a co-citation network. The scientific basis of this reverse recommendation path is based on the following: First, co-citation is a recognized knowledge association measurement method in bibliometrics. Papers that share the same references have inherent similarities in research topics and methodologies. Therefore, the paper groups in the co-citation network often submit to the same or similar journal groups. Second, the scope of academic journals and their preferences for manuscript sources are stable. Highly similar papers in the same research direction tend to be published in a few related journals. Therefore, the actual publication destination of similar papers can serve as an objective reference for journals suitable for this research direction. Third, this channel serves as an auxiliary evidence dimension (S_evidence) in the five-dimensional scoring, working in conjunction with topic matching (S_topic) and subject matching (S_subject). It is not an independent decision-making basis, but rather provides supplementary verification signals from actual published data for the positive matching channel.

[0029] Step S7: Candidate Set Consolidation and Deduplication. The system uses the journal's unique identifier as the deduplication key to deduplicate the recall results from the three channels: exact topic matching, fuzzy subject matching, and co-citation network enhancement. For journals recalled by multiple channels simultaneously, the highest topic matching score is retained as the ranking criterion. After merging, the journals are sorted by topic matching score to obtain the candidate journal set. The maximum number of journals in the candidate set is 100.

[0030] Step S8: Five-Dimensional Weighted Scoring Stage. The system performs a five-dimensional weighted scoring on the candidate journals, calculating the comprehensive matching score (Score). The scoring formula is: Score = w1 x S_topic + w2 x S_subject + w3 x S_impact + w4 x S_cost + w5 x S_evidence. The weight coefficients satisfy the priority relationship w1>w2 ≥ w5>w3>w4 (in this embodiment, the values ​​are w1=0.35, w2=0.20, w3=0.15, w4=0.10, w5=0.20). The meanings of each dimension are as follows: S_topic (topic matching degree, weight 35%): Calculated based on the overlap between the paper's keywords and OpenAlex Topics, reflecting the degree of matching between the paper's topic and the journal's coverage. S_subject (subject matching degree, weight 20%): Calculated based on the degree of match between the subject of the paper and the scope of journal coverage, reflecting the degree of matching at the subject level; S_impact (impact matching, weight 15%): Calculated based on whether the journal's impact factor and h-index are within the user's expected range, reflecting the degree of matching between the journal's impact and the user's expectations; S_cost (peer review cost matching, weight 10%): Calculated based on whether the peer review cycle and OA fee meet the user's requirements, reflecting the degree of matching between the submission cost and the user's expectations; S_evidence (paper publication evidence, weight 20%): Calculated based on the number of publications and citations of similar papers obtained from the co-citation network enhancement channel in step S6 in each candidate journal. It reflects the credibility of the actual publication destination of published papers as a basis for recommendation. This dimension is one of the core innovations of this invention, which uses the actual publication destination of similar published papers as evidence for recommendation.

[0031] Step S9: Risk Assessment Phase. The system conducts a multi-dimensional risk assessment of the recommendation results, including: Risk grading of the proportion of publications by Chinese authors: Based on the past 5 years of publication data by Chinese authors from OpenAlex, the system calculates the proportion of publications by Chinese authors. The core principle of this grading method is to use the natural proportion of Chinese authors in global academic output as a baseline. When the proportion of Chinese authors in a journal deviates significantly from this baseline, it indicates that the journal's manuscript source structure is abnormal and there may be quality risks. According to the NSF's "Science & Engineering Indicators" report and statistics from the OpenAlex global paper database, the proportion of Chinese authors in global academic papers is approximately 20-25%. The system uses this as a baseline and grades the risk according to the following rules: A proportion below the baseline (approximately 25%) is marked as safe, indicating that the journal's author sources are diversified; a proportion between the baseline and twice the baseline (approximately 25%-50%) is marked as friendly to Chinese authors; a proportion between twice and 2.8 times the baseline (approximately 50%-70%) is marked as high risk. Level 1: The percentage of Chinese authors publishing between 2.8 and 3.4 times the baseline (approximately 70%-85%) is marked as high-risk; the percentage exceeding 3.4 times the baseline (approximately 85%) is marked as extremely high-risk. The above classification thresholds are determined based on the percentage of publications by Chinese authors globally (approximately 25%), with classification nodes set according to a proportionally increasing principle. This classification method references the assessment approach regarding author source concentration in the journal warning list released by the Documentation and Information Center of the Chinese Academy of Sciences and New Academic Journals, as well as the DOAJ's evaluation criteria for the internationalization of journals.

[0032] OA type standardization: The system standardizes the OA types returned by OpenAlex, including top journal coverage rules (forcing top journals such as Nature, Science, and Cell to be marked as hybrid type), non-standard type fallback mapping (green is mapped to closed, bronze is mapped to hybrid), and DOAJ inclusion status inference (gold and diamond types are marked as included, closed and hybrid types are marked as not included).

[0033] Warning Journal Marker: The system checks whether a journal is on the warning journal list, and if so, adds a warning mark.

[0034] Step S10: Recommendation Reason Generation and Result Output Stage. Based on scores across various dimensions and evidence from the data source, the system generates fact-based recommendation reasons for each recommended journal. The recommendation reasons cite specific numerical values ​​from the data source (such as the number of topic matches, the number of similar paper evidences, and risk indicator values) to improve the interpretability of the recommendations. The system returns a Top-N recommendation result (default Top 5) with matching scores, recommendation reasons, and risk warnings.

[0035] Example 2: This example provides an intelligent recommendation system for academic journals based on multi-source heterogeneous data fusion, comprising the following six-layer architecture: (1) User input layer, which is responsible for receiving the user's paper information and preference settings. The paper information includes the title, abstract and keywords, and the preference settings include the expected impact factor range, expected review period, expected databases, etc. (2) Feature extraction layer: Based on the large language model API, the paper text is intelligently analyzed to extract structured features, including keyword extraction, subject classification and research type identification; (3) Multi-source data fusion layer, including OpenAlex data source module, Crossref data source module, SemanticScholar data source module and unified data standardization module. The three data source modules are called in parallel and asynchronously. If any module fails to call, it will not block other modules. The unified data standardization module converts the heterogeneous data returned by the three data sources into the internal standard schema. (4) Three-channel recall layer, including topic exact matching channel, subject fuzzy matching channel and co-citation network enhancement channel, as well as candidate set and deduplication module. The co-citation network enhancement channel selects seed papers from semantic search results, expands the co-citation network using the relatedWorks field of OpenAlex, counts the publication frequency of related papers in journals, and realizes the reverse recommendation path from paper to journal. The three channels are executed in parallel, and after merging and deduplication, they are sorted according to topic matching degree. (5) The scoring and ranking layer includes a five-dimensional weighted scoring module, a risk assessment module, and a recommendation reason generation module. The five-dimensional weighted scoring module calculates a comprehensive matching score based on five dimensions: topic matching degree, discipline matching degree, influence matching degree, peer review fee matching degree, and paper publication evidence degree. Among them, the paper publication evidence degree is calculated based on similar paper publication data obtained through the co-citation network enhancement channel. The risk assessment module includes a risk classification sub-module for the proportion of Chinese authors' publications, an OA type standardization sub-module, and a warning journal marking sub-module. The recommendation reason generation module generates recommendation reasons based on evidence citation. (6) Recommendation output layer, which outputs Top-N recommendation results with matching scores, recommendation reasons and risk warnings.

[0036] Example 3: Application Scenario Example. A researcher has completed a paper titled "Deep Learning for Medical Image Analysis: A Multi-Modal Approach" and needs to select a suitable SCI journal for submission. The researcher expects an impact factor between 3 and 8 and a review period of no more than 3 months.

[0037] System Flow: (1) The user enters the paper title, abstract, and keywords; (2) The system calls AI services to extract paper features: the main discipline is "computer science", the minor disciplines are "medical imaging" and "deep learning", and the research type is "experimental research"; (3) The system executes three recall channels in parallel: OpenAlex semantic search returns 20 relevant papers, counts the frequency of publication journals, topic matching recalls 30 candidate journals, and subject matching recalls 50 candidate journals; (4) The co-citation network has been expanded to include 8 new candidate journals; (5) After merging and deduplicating, 65 candidate journals were obtained; (6) After filtering by IF range (3-8), 35 books remain; (7) Batch acquisition of Crossref and OpenAlex augmented data; (8) Five-dimensional weighted scoring, select the Top 5; (9) Risk assessment: One book has 72% Chinese authors (high risk), and a risk warning has been added; (10) Output recommendation results, with each journal accompanied by a matching score, recommendation reason and risk warning.

[0038] Technical performance: Recommendation accuracy (user satisfaction rating) is greater than or equal to 80%; recommendation response time is less than 10 seconds; recommendation coverage (recommendation coverage of journals in different disciplines) is greater than or equal to 80%.

Claims

1. An intelligent recommendation method for academic journals based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: S1: Receive the user's paper information and preference settings, where the paper information includes the paper title, abstract and keywords, and the preference settings include the expected impact factor range, expected review period and expected databases; S2: Call the large language model API to extract features from the paper text and obtain structured paper features, which include keyword list, main subject classification, sub-subject direction and research type; S3: Based on the structured paper features obtained in step S2, perform multi-source heterogeneous data collection in parallel, call multiple academic open data source interfaces to obtain journal metadata, paper statistics and semantic search results respectively, and adopt a parallel asynchronous mode so that the failure of any data source call does not block the call of other data sources. S4: Convert heterogeneous data returned from various data sources into an internal standard schema through a unified data standardization module to achieve unified field mapping and data structure; S5: Parallel execution of dual-channel candidate journal recall, including: S5a: The topic precise matching channel, based on the subject topic map, calculates the overlap between the paper keywords and the journal topic, and recalls candidate journals by sorting the number of overlaps; S5b: Subject-based fuzzy matching channel, which performs fuzzy search based on multiple sets of mainstream subject classification fields to recall candidate journals; S6: Select seed papers from semantic search results, expand the co-citation network using the literature association field, obtain relevant papers and count their publication journal frequency, add high-frequency journals to the candidate set as a co-citation network enhancement channel, and realize the reverse recommendation path from paper to journal. S7: Merge and deduplicate the recall results from the three channels in steps S5 and S6, and sort them by topic matching degree to obtain a set of candidate journals; S8: Perform a five-dimensional weighted scoring on the candidate journals and calculate the comprehensive matching score (Score). The five-dimensional weighted scoring formula is as follows: Score = w1 × S_topic + w2 × S_subject + w3 × S_impact + w4 × S_cost +w5 × S_evidence Where S_topic is the topic matching degree, S_subject is the subject matching degree, S_impact is the impact matching degree, S_cost is the peer review fee matching degree, S_evidence is the paper publication evidence degree, w1 to w5 are the corresponding weight coefficients, and S_evidence is calculated based on the number of publications and citations of similar papers obtained from the co-citation network enhancement channel in step S6 in each candidate journal. S9: Conduct multi-dimensional risk assessments on the recommendation results, including risk grading based on the proportion of articles published by Chinese authors, standardization of OA types, and marking of warning journals, and outputting structured risk warnings; S10: Based on scores across various dimensions and evidence from data sources, generate fact-based recommendation reasons for each recommended journal and return Top-N recommendation results with matching scores, recommendation reasons, and risk warnings.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data acquisition mentioned in step S3 specifically includes: Retrieve journal topic graphs, h-index, i10-index, OA status, APC fees, and publication statistics of Chinese authors from the OpenAlex API; Retrieve journal publication volume, citation statistics, annual trends, and paper type distribution data from the Crossref API; Obtain semantic search results, TLDR summaries, high-impact citation tags, and similar paper recommendation data from the Semantic Scholar API.

3. The method according to claim 1, characterized in that, The weighting coefficients mentioned in step S8 satisfy the following priority relationship: w1 > w2 ≥ w5 > w3 > w4, where w1 is the topic matching weight, which is the maximum value among the five dimensions of scoring; w2 is the subject matching weight; w5 is the paper publication evidence weight; w3 is the influence matching weight; and w4 is the peer review fee matching weight, which is the minimum value. The priority relationship of the weighting coefficients is determined based on the degree of influence of each decision dimension on the submission decision in the journal recommendation scenario. The topic matching has the greatest impact on the submission decision, while the peer review fee has the least impact.

4. The method according to claim 1, characterized in that, The risk classification based on the proportion of publications by Chinese authors mentioned in step S9 specifically refers to: We analyzed data on articles published by Chinese authors over the past five years and calculated the percentage of articles published by Chinese authors. A baseline is determined based on the proportion of natural publications by Chinese authors in global academic output. When the proportion is lower than the baseline, it is marked as a safe level. When the percentage is higher than the baseline but does not exceed the preset threshold for friendliness to Chinese people, it is marked as a friendliness level to Chinese people; When the percentage exceeds the preset threshold for friendliness towards Chinese people, it is marked as a high-risk level; When the proportion exceeds the preset high-risk threshold, it is marked as high-risk. When the proportion exceeds the preset extremely high risk threshold, it is marked as an extremely high risk level; The baseline mentioned above is determined based on the historical publication percentage of Chinese authors in global academic paper databases.

5. The method according to claim 4, characterized in that, The baseline is the statistical average of the percentage of papers published by Chinese authors in the global academic paper database over the past 5 years; the "Chinese-friendly" threshold is 1.5 to 2 times the baseline; the "high-risk" threshold is 2.5 to 3 times the baseline; and the "extremely high-risk" threshold is 3 to 3.5 times the baseline.

6. The method according to claim 1, characterized in that, The OA type standardization process described in step S9 specifically includes: Force Nature, Science, and Cell journals to be labeled as hybrid type; Perform fallback mapping for non-standard OA types, where green is mapped to closed and bronze is mapped to hybrid; The inclusion status of DOAJ is inferred based on the OA type, where the gold and diamond types are marked as included, and the closed and hybrid types are marked as not included.

7. The method according to claim 1, characterized in that, The selection of seed papers from semantic search results in step S6 specifically involves selecting the top N papers with the highest relevance in the semantic search results as seed papers, where N ranges from 2 to 5. The expansion of the co-citation network specifically involves obtaining the relatedWorks list for each seed paper, batch obtaining the metadata of related papers, counting the number of related papers included in each journal and the cumulative number of citations, and adding journals with more than a preset threshold to the candidate set.

8. The method according to claim 1, characterized in that, The merging and deduplication process described in step S7 is as follows: using the journal's unique identifier as the deduplication key, the recall results from the three channels are deduplicated; for journals that are recalled by multiple channels at the same time, their highest topic matching degree is retained as the sorting criterion; the upper limit of the candidate set is 100 journals.

9. An intelligent recommendation system for academic journals based on multi-source heterogeneous data fusion, characterized in that, include: The user input layer is used to receive users' paper information and preference settings; The feature extraction layer, connected to the user input layer, extracts features from the paper text based on the large language model API to obtain structured paper features; The multi-source data fusion layer includes several academic open data source modules and a unified data standardization module. The unified data standardization module is used to convert the heterogeneous data returned by each data source module into an internal standard schema. The three-channel recall layer includes a topic precise matching channel, a subject fuzzy matching channel, and a co-citation network enhancement channel, as well as a candidate set and deduplication module for merging and deduplicating the recall results of the three channels. The co-citation network enhancement channel selects seed papers from semantic search results, expands the co-citation network using document association fields, and counts the publication frequency of related papers in journals to realize a reverse recommendation path from papers to journals. The scoring and ranking layer includes a five-dimensional weighted scoring module, a risk assessment module, and a recommendation reason generation module. The five-dimensional weighted scoring module calculates a comprehensive matching score based on five dimensions: topic matching degree, discipline matching degree, influence matching degree, peer review fee matching degree, and paper publication evidence degree. The paper publication evidence degree is calculated based on similar paper publication data obtained through the co-citation network enhancement channel. The recommended output layer is used to output Top-N recommendation results with matching scores, recommendation reasons, and risk warnings.

10. The system according to claim 9, characterized in that, The risk assessment module includes: The risk classification submodule for the proportion of publications by Chinese authors is used to calculate the proportion of publications by Chinese authors based on the publication data of Chinese authors over the past 5 years from OpenAlex statistics, and to determine the baseline based on the natural proportion of publications by Chinese authors in global academic output, and to classify the risk according to the preset multi-level thresholds. The OA type standardization submodule is used to standardize the OA types returned by OpenAlex, including top journal coverage rules, non-standard type fallback mapping, and DOAJ inclusion status inference. The Warning Journal Tagging submodule is used to check whether a journal is in the warning journal list and add a tag.

Images