Interdisciplinary scientific research potential assessment method based on dynamic multi-modal knowledge graph
Through the interdisciplinary scientific research potential assessment method based on dynamic multimodal knowledge graph, new academic concepts and sudden hot spots are identified in real time, and the problem of difficulty in taking into account update speed and accuracy in the existing technology is solved, and the timeliness and accuracy of universities in the strategic deployment of disciplines and resource allocation is achieved.
Patent Information
- Application Number
- CN202510665367.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-22
- Publication Date
- 2025-06-20
- Estimated Expiration
- 2045-05-22
AI Technical Summary
It is difficult for existing technologies to accurately identify new academic concepts and mutation hot spots while ensuring the update speed, resulting in lag or mismatch between universities in the strategic deployment of cutting-edge disciplines and resource allocation.
The interdisciplinary scientific research potential assessment method based on dynamic multimodal knowledge graph is adopted, and the map structure is updated in real time through concept drift detection and ontology expansion, and the emerging hot spots are identified in combination with timing difference indicators such as citation growth rate and cooperative networks, and the accuracy of the evaluation model and continuous iteration ability are maintained with the help of online learning and cross-domain migration strategies.
Real-time capture of new academic concepts and sudden hot spots has been achieved, helping universities to track the pulsation of discipline evolution, optimize resource allocation, improve perceived accuracy of emerging research fields, and ensure the timeliness and accuracy of decision-making information.
Smart Images

Figure CN120181409A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of scientific research management and intelligence analysis, and specifically to an interdisciplinary scientific research potential evaluation method based on a dynamic multi-modal knowledge graph. Background Art
[0002] In today's university scientific research environment, with the continuous advancement of digitalization and informatization, a large number of academic achievements emerge in various forms such as papers, patents, reports, datasets, etc. Researchers also conduct cross-field exchanges and knowledge sharing through diverse channels such as online academic conferences, social media, and video forums. Especially during the rapid rise of emerging disciplines and interdisciplinary fields, changes in funding support, orientation preferences, and social needs can lead to frequent evolutions in the cooperation networks and topic layouts among different research institutions. For example, some laboratories may achieve breakthroughs in the combination of biology and information engineering, or suddenly appear interdisciplinary research projects targeting public health emergencies; these dynamic factors make the scientific research ecosystem highly non-linear and decentralized. If relying solely on traditional text mining and annual aggregated literature databases to track the development of disciplines, it is often impossible to promptly capture emerging hotspots and difficult to reflect the true context of scientific research activities in a timely manner; this not only affects the strategic deployment of universities in frontier disciplines but may also result in missed funding opportunities or technology cooperation opportunities.
[0003] Therefore, the industry urgently needs a technical solution that can perform incremental acquisition and comprehensive processing on different sources (text, video, patent descriptions, etc.), realize real-time discovery and forward-looking evaluation of interdisciplinary potential, and accurately depict the new dynamics in the scientific research field under the drive of a complex social environment and multiple guiding factors.
[0004] However, when faced with these multi-source heterogeneous data, how to balance the accurate identification of new academic concepts and mutation hotspots while ensuring the update speed has become a core technical problem. Usually, information systems need to transform a large amount of academic data in different modalities into computable knowledge structures and track the evolution of disciplinary themes or collaboration relationships over time. However, due to the irregular update cycle of scientific research information and the possible explosive growth or rapid decline of cross-field cooperation in the short term, traditional once-built knowledge graphs or static evaluation models are difficult to timely warn of these rapidly changing research trends; when the administrators of research institutions or universities have not yet perceived the potential of interdisciplinary research, the rapidly changing disciplinary landscape has already been reorganized, resulting in lags or mismatches in resource allocation, disciplinary support, talent introduction, etc. Without an effective mechanism for integrating dynamic incremental data and a sensitive detection of sudden trends, the constructed evaluation system will be disconnected from the real scientific research ecosystem, ultimately leading to problems such as distorted decision-making information, uneven funding distribution, and damaged competitiveness. Therefore, how to perform multi-modal and incremental data collection and analysis on cross-field research activities in a short time and map this dynamic change to scientific research potential evaluation in real time has become a key technical problem that urgently needs to be solved. Summary of the Invention
[0005] (I) Technical Problems to be Solved Aiming at the deficiencies of the prior art, the present invention provides a cross-disciplinary scientific research potential evaluation method based on a dynamic multi-modal knowledge graph. By detecting concept drift and expanding the ontology, the graph structure is updated in real time. Combining time series differential indicators such as citation growth rate and cooperation network to identify emerging hotspots; and using online learning and cross-domain transfer strategies to maintain the accuracy and continuous iteration ability of the evaluation model in multiple environments. Finally, through scenario simulation, the potential evolution is quantified and strategic suggestions are output, supplemented by multi-modal embedding to further enhance the sensitivity to cross-disciplinary fields, forming a closed-loop evaluation system from data to decision-making, helping universities with interdisciplinary layout and resource allocation, and solving the technical problems described in the background art.
[0006] (II) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: A cross-disciplinary scientific research potential evaluation method based on a dynamic multi-modal knowledge graph, including that if the amount of newly crawled academic data exceeds the threshold, text mining is used to extract entities and relationships, and after identifying emerging terms through concept drift detection, they are incrementally merged into the incremental knowledge graph, and then the ontology level is expanded to accommodate the new concepts; After reaching a new time slice, calculate the multi-time window differential vector and aggregate it to highlight the theme growth or convergence trend, generate a cross-trend indicator and a list of hot topics and mark sudden rises and falls; When the cross-trend indicator and the hot topic list show significant fluctuations within the domain, online parameter updates are performed to optimize the model parameters, and the migration loss function is used to adaptively fine-tune the model parameters in different regions; When the optimized model is obtained by solving and the management terminal submits a scenario composed of external environment parameters and scientific research orientation variables, the scenario potential function is called to evaluate the potential trend of resource allocation and interdisciplinary collaboration, output comparable indicators, and when the actual and predicted differences are greater than expected, a deviation signal is fed back to the online learning module; Perform modal marking and semantic vector extraction on the generated multi-modal original dataset. The aggregation function aligns cross-modal features and outputs a multi-modal fusion library, which is incrementally incorporated into the multi-modal knowledge graph, and triggers re-fine-tuning of the optimized model when anomalies occur.
[0007] Furthermore, after establishing a timed or real-time crawling strategy for multi-source data, set the crawling frequency and filtering conditions to obtain a phased original dataset; use text mining algorithms to perform entity annotation on the common structures in the academic fields of the phased original dataset, and perform preliminary merging of duplicate entities or name variants to obtain an entity set. After extracting the association information between documents, form a preliminary relationship set.
[0008] Furthermore, incrementally merge the entity set and the relationship set into the incremental knowledge graph; if overlap with existing nodes or attribute conflicts are detected, adopt a merging strategy based on a confidence threshold to correct them; perform concept drift calculation for nodes and edges in the incremental knowledge graph, If the obtained concept drift measure exceeds the preset drift threshold, use it as a new concept node and trigger corresponding ontology expansion or hot topic marking processing; incorporate uncollected interdisciplinary fields into the semantic structure of the graph and mark them as hot topic candidates, and output a hot topic candidate list.
[0009] Furthermore, add marks to the hot topic candidate list so that it includes an additional cross-potential dimension when generating dynamic feature vectors; based on the incremental knowledge graph, aggregate and generate serialized graph slices, extract key information from each graph slice, and generate feature vectors of nodes through embedding learning; extract the difference or high-order change rate of the feature vectors of adjacent graph slices to form a time-series difference vector.
[0010] Furthermore, for each topic node, calculate an aggregation function that measures the overall evolution trend across time periods; when the trend aggregation function exceeds the preset aggregation threshold, determine the corresponding topic as a node with significant cross-potential and mark it as a cross-trend indicator, and summarize it into the global trend information at the entire graph dimension; combine the time-series difference vector and the cross-trend indicator to screen the detected new concept nodes and obtain a hot topic list.
[0011] Further, after obtaining new academic information and combining it with the hot topic list, select nodes and relationships with a high degree of relevance to the new hot topics in the incremental knowledge graph and construct an incremental training set. Combine the incremental training set and cross-trend indicators to construct an online update loss function, and perform iterative solution on the online update loss function to obtain an online update model after updating the model parameters.
[0012] Further, by analyzing the node distribution and external environment labels in the incremental knowledge graph, distinguish the source domain and the target domain, and define a transfer loss function to characterize the distribution difference between the source environment and the target environment. If the transfer loss function is greater than the expected value, based on the existing online update model in the source environment, perform transfer optimization on the incremental data in the target environment, and perform gradient solution on the defined comprehensive objective function to obtain an optimized model.
[0013] Further, after collecting the external environment parameters and scientific research orientation variables, associate them with the topic nodes or cooperation networks in the incremental knowledge graph; set the change nodes of the environment or orientation in segments within the simulation duration to form a scenario function that evolves over time for dynamic loading of scenario configurations; finally, output a scenario definition set, where each scenario corresponds to a set of environment-orientation combinations and their evolution details in the time dimension.
[0014] Further, send the updated node embedding feature vector representation into the optimized model to output a new cross-potential prediction value; the simulation-driven function calculates the potential in each sub-period in combination with the scenario configuration and then accumulates it to the next period until the entire simulation duration is covered; calculate the scenario potential for all scenarios in the scenario definition set respectively, and then perform sorting or classification summary to obtain a multi-scenario potential comparison.
[0015] Further, define a policy benefit function in combination with actual management requirements to select scenarios. When there are significant differences in scenario potential or obvious deviations from the expected value, collect the deviation information and external feedback during the simulation process into a new batch of incremental data, perform model correction or weighted reinforcement, and output a policy plan and a risk list, including the priority policy combination and response plans under different environmental assumptions.
[0016] Further, after introducing multi-modal information, incorporate it into the data collection pipeline, use their respective corresponding collection strategies to generate a multi-modal raw data set; perform preliminary modal differentiation on the multi-modal raw data set, extract core metadata, and then perform semantic annotation to generate multi-modal labeled data. If the same event appears in multiple modalities, align it through multiple factors and merge it into the same entity or relationship node.
[0017] Furthermore, feature vectors are extracted from the data segments of each modality using a deep learning model to form a temporary multi-modal vector set; a cross-modal alignment transformation is introduced to map the features of different modalities and aggregate them in a unified representation space, and a multi-modal semantic vector is output; the multi-modal semantic vector is cross-checked with the keyword tags and hot topic candidate lists generated by text mining, and if abnormal noise is found, a concept drift or inconsistency detection mechanism is enabled for correction.
[0018] Furthermore, corresponding nodes or relationships in the incremental knowledge graph are created or updated for each newly emerging or significantly changed entity, and an online learning and transfer learning and scenario simulation and strategy verification mechanism is introduced for rapid verification. If the multi-modal information indicates that the frequency of occurrence of a certain hot topic surges in videos or experimental logs, fine-tuning of the optimized model is triggered; if the collaborative research methods in a certain field are more intensive, the corresponding potential value or investment ratio is increased in the simulation session; when the multi-modal update causes changes to the interdisciplinary trend or hot topic list, the corresponding changes and potential deviation information are sent back again.
[0019] (III) Beneficial Effects The present invention provides an interdisciplinary scientific research potential evaluation method based on a dynamic multi-modal knowledge graph, having the following beneficial effects: Through the incremental knowledge graph and the concept drift detection strategy, real-time capture of new academic concepts and sudden hot topics is realized, avoiding the situation of missing important research directions due to the obsolescence of the knowledge base in traditional static evaluations; Time series analysis focuses on differential indicators such as keywords, cooperation networks, and citation growth rates on different time slices, enabling cross-topics that are rapidly emerging to be identified and quantified to output a hot topic list for the first time , which not only helps university management to track the pulse of disciplinary evolution but also lays a data foundation for subsequent online model fine-tuning; Online learning and transfer learning endow the model with adaptability: when a significant change in the disciplinary cooperation pattern or external environment is detected, the evaluation parameters are fine-tuned or cross-domain transferred according to the incremental data or distribution differences, taking into account the original knowledge accumulation and adaptation to new trends, being able to respond faster to short-term academic fluctuations, retaining the long-term experience of the algorithm and avoiding evaluation distortion; Scenario simulation and strategy verification introduce multi-scenario configurations and scenario potential functions to evaluate the potential trends of interdisciplinary disciplines under different orientation and environmental combinations, helping the management side to intuitively compare multiple decision-making schemes and quickly select the best balance point. If a deviation is found, the result can also be fed back to the online learning module for self-correction, forming a deep collaboration of data - model - strategy; Multi-modal data fusion is achieved through multi-modal vectors such as images, audio, conference videos, etc. Combined with the incremental knowledge graph, capture the subtle clues of interdisciplinary interactions and incorporate them into the real-time judgment of hot topics, significantly improving the perception accuracy of emerging research fields; The above elements interact with each other on the core of the incremental knowledge graph: the data collection link provides fresh information, the time series trend analysis provides hot spot positioning, online and transfer learning maintain the model's agile adaptation to the changing environment, scenario simulation transforms the above results into actionable decision support, and multimodal fusion further broadens the information source. Through the synergistic effect of these multi-links, taking into account the continuity of data management and model training, it helps universities or research institutions to timely grasp cross-disciplinary opportunities, optimize resource allocation and maintain long-term competitiveness in a dynamic and multi-source academic environment. Brief Description of the Drawings
[0020] Figure 1 It is a schematic flowchart of the method for evaluating the interdisciplinary scientific research potential based on the dynamic multimodal knowledge graph of the present invention. Detailed Embodiment
[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0022] Please refer to Figure 1 , the present invention provides a method for evaluating the interdisciplinary scientific research potential based on the dynamic multimodal knowledge graph, including, Step 1. When it is monitored that the amount of new data such as academic literature, patents and scientific research news exceeds the threshold, extract the entity set and the relationship set through a text mining algorithm, then call the concept drift detection function to expand the ontology of high-growth concepts, and perform incremental merging of new nodes and associations into the incremental knowledge graph based on the semantic hierarchical structure to realize the dynamic update of the disciplinary structure and the real-time capture of cross-disciplinary hotspots; The content of the above Step 1 includes the following: Step 101. Construction of a multi-source data collection pipeline and preprocessing of entity associations Configure a unified collection module, establish a timed or real-time crawling strategy for multi-source data such as academic papers, patent documents, scientific research news, and guiding documents, and obtain a phased raw data set after setting the crawling frequency and filtering conditions (such as specifying topics, keywords or institutional sources) , where represents the collection time or time batch; Use a text mining algorithm for the phased raw data set Perform batch processing, perform entity annotation for common structures in the academic field (such as author names, institutional names, professional term topic concepts, etc.). To reduce noise, it is necessary to initially merge duplicate entities or name variants to obtain an entity set ; Extract the association information between documents in terms of citation, author cooperation, keyword co-occurrence, etc., to form a preliminary relationship set , and these entities and relationships are temporarily stored as temporary structured data ; Perform unique identifier ID assignment on the newly collected data, and record the phased original data set as the input and the mapping relationship between the temporary structured data as the output to ensure accurate traceability and invocation in subsequent steps; When in use, multi-source data is centrally managed through the same pipeline, ensuring that subsequent steps do not need to repeat scraping or cleaning, reducing system resource waste. Assign unique IDs to all newly identified entities and relationships and record their time batches to achieve traceability and consistency in subsequent steps.
[0023] Step 102, Incremental Knowledge Graph Generation and Ontology Extension After receiving the temporary structured data , merge the entity set and the relationship set incrementally into the incremental knowledge graph ; if overlap with existing nodes or attribute conflicts are detected, a merging strategy based on a confidence threshold is adopted for correction. Specifically, calculate the comprehensive score based on name similarity and cosine similarity of embedding vectors. If the comprehensive score exceeds the confidence interval, it is judged as a duplicate node, and merge is performed according to the frequency of text attributes first and the credibility-weighted average method for numerical attributes. If it is below the confidence interval, a new node is created; when it is between the two, it is marked as a conflict pending review and finally confirmed by manual or semi-automatic processes; the incremental knowledge graph generated in this process is the latest state of the time-oriented knowledge graph and serves as the benchmark for subsequent time series analysis; To timely identify new or significantly changed academic concepts, it is necessary to perform concept drift calculation on nodes (entities) and edges (relationships) in the incremental knowledge graph , where: To measure the dynamic changes in the embedding space of the knowledge graph within the time period , define the concept drift measure function as:
[0024] where: is the moment The incremental knowledge graph below contains a node set and its associated relationship information; is the set of valid node (entity) indices at time ; is the -dimensional vector embedding value of the node at continuous time ; if there is only embedding data at discrete time points, it can be constructed by linear or high-order interpolation ; is the partial derivative or difference approximation of the above dimensional embedding value with respect to time , reflecting the rate of change of the node embedding in the dimension over time; is the weight factor for the dimension, with a value between 0 and 1, used to strengthen or weaken certain feature dimensions; is a high-order distance exponent greater than 1; is the time interval; If the obtained concept drift measure exceeds the preset drift threshold , it can be determined that significant semantic evolution has occurred, significant concept drift has occurred, and it is used as a new concept node, which can then trigger corresponding ontology expansion or hot spot marking processing to automatically update the classification hierarchy of newly emerged or significantly changed concept nodes; Incorporate potentially unrecorded interdisciplinary fields into the semantic structure of the knowledge graph, mark these newly added or updated nodes as hot spot candidates, and output a list of hot spot candidates ; When in use, ensure that the knowledge graph can evolve rapidly with data iteration through incremental merging and present it in a unified structural form, providing reliable continuity for subsequent time series analysis. Based on the detection mechanism of the concept drift measure , automatically identify new hot spot concepts, reduce the cost of manual maintenance of the ontology hierarchy, and avoid missing potential important interdisciplinary fields; combine entity embedding with the change amplitude to not only capture new nodes but also quantify the semantic drift of existing concepts to achieve early warning of subtle semantic changes.
[0025] Step 2. When the incremental knowledge graph is updated and reaches the next time slice , based on the differential mapping and combined with indicators such as citation growth rate and cooperation frequency, perform time series slicing and trend aggregation on the discipline theme and cooperation network, and output the cross-trend indicator and the list of hot spot themes , and associate and verify them with the concept drift trajectory; Step 2 includes the following contents: Step 201, Time Slicing and Dynamic Feature Extraction Based on the incremental knowledge graph aggregate the data storage and marking of different time batches, and generate serialized graph slices according to the set time window (such as quarterly, annual or adaptive time window) , denoted as:
[0026] where represents the timestamp or time interval index of each time period; these time-sequential graphs retain the incremental update information of nodes and relationships after ontology expansion, and can accurately record the evolution status of subject concepts and cooperation relationships in different time periods.
[0027] For each graph slice extract key information such as subject themes, cooperation networks, citation relationships, etc., and generate feature vectors of nodes through embedding learning methods (such as embedding models based on graph representation learning or text representation learning) where represents the node identifier existing at time , contains multi-dimensional encodings such as topic labels, cooperation frequencies, citation relationships, etc.; To identify interdisciplinary intersections, it is necessary to add marks to the hot candidate list so that it includes an additional cross-potential dimension when generating dynamic feature vectors; the higher the value of this dimension, the more neighbors of different disciplines the node connects in the existing cooperation network, and its semantic embedding has undergone significant drift; this cross-potential dimension can be combined with the results of concept drift measurement in step 102 (such as a large number of newly added concept nodes) for enhanced encoding, facilitating focused attention in subsequent trend detection; Extract the difference or high-order change rate of the feature vectors of adjacent graph slices to form a time-series difference vector , and this difference value can help judge the increase or decrease trend of a certain subject theme or cooperation relationship between different time periods, avoiding ignoring potential clues of sudden increase or decrease only based on absolute values.
[0028] Through time slicing and feature difference between slices, while maintaining the historical data of nodes, dynamic changes can be accurately captured. After marking the hot candidate list , the evolution of emerging interdisciplinary entities can be more directly concerned in subsequent analysis.
[0029] Step 202, Trend Aggregation and Cross-Hotspot Detection To measure the overall evolution trend across time periods, a multi-time-window trend aggregation function is defined and denoted as:
[0030] where: , representing a certain theme or node at different time points obtained vectorized representations (such as integrating information on cooperation frequency, citation characteristics, interdisciplinary labels, etc.); is an ordered set of time slices or endpoints of time intervals; is the theme or node at time partial derivative of the vectorized representation with respect to (if there are only discrete vectors, it can be approximated by taking differences after interpolation); is a set of matrix weights that change over time and are used to further correct or amplify the changes in node vectors during integration. It can be regarded as a dynamic mapping for factors such as external research orientation, funding, or interdisciplinary degree; is a user-defined high-order or weighted vector norm, such as norm ; is the exponent used to amplify the result of this norm, ; is the weight coefficient of the time window and its value is greater than or equal to 0; For each theme node and , calculate the trend aggregation function over each time period , that is, obtain the cumulative change value of the theme . When the trend aggregation function exceeds the preset aggregation threshold, it is determined that there is a significant anomaly in the research activity, citation frequency, or cooperation scale of this theme, and it is a node with significant cross-potential, marked as the cross-trend indicator , and this indicator will be summarized into the global trend information of the entire map dimension and finally output as ; Combining the high-order difference data generated in step 201 and the cross-trend indicator , entities that are detected as newly emerging for the first time or whose embedded representations have a significant semantic drift from history, that is, new concept nodes, are screened out with emphasis.
[0031] If some new concept nodes or nodes with significant cross-potential show a sudden increase in a short period of time and have a relatively high degree of association across multiple disciplines, they will be included in the hot topic list ; Ultimately, step 202 will produce two key results: Cross Trend Indicator : Measures the evolution of different topic nodes within a given time interval; Hot topic list : Focus on emerging disciplines and topics with high growth rates or high interdisciplinary relevance; When using it, you can take into account the importance of different time windows to ensure that both rapidly emerging hot spots and areas of sustained and stable development can be reasonably evaluated; highlight potential cross-explosion points, amplify and identify those topics with extremely high rates of change in a short period of time, and improve the sensitivity of capturing explosive cross-disciplinary points. Not only can it output quantitative cross-trend indicators , and can also provide an additional list of hot topics , providing an intuitive reference for subsequent model optimization and strategy formulation.
[0032] First, we need to map the incremental knowledge graph on the timeline. Slicing is performed and dynamic feature vectors are constructed. Then, the multi-time window trend aggregation formula and high-order norm measurement are used to identify the intersections and hot spots of emerging disciplines. The output cross-trend indicators and hot topic lists can not only provide accurate model update basis for the third step of online and transfer learning model optimization, but also lay the foundation for trend judgment and strategy evaluation in the fourth step and final scenario simulation stage.
[0033] Step 3: Get the cross trend indicator and Hot List When significant changes in the scientific research environment are detected, the incremental data Perform weighted update , and use the migration loss function Adaptively fine-tune model parameters for different regions or universities to continuously maintain evaluation accuracy and environmental matching; The step three includes the following contents: Step 301: Online learning and model parameter fine-tuning Receive cross trend indicators , Hot Topic List And incremental knowledge graph After that, the original evaluation model parameters obtained from the previous round of training are retained. , used for online learning initialization to avoid training from scratch; When new academic information (papers, patents, scientific research news, etc.) arrives, combine it with the hot topic list , in the incremental knowledge graph Select nodes and relationships with high correlation with these new hot topics to build an incremental training set , the incremental training set Focus on the subject themes that have appeared frequently recently or have high CTI values to strengthen the adaptation to new trends; Combine the incremental training set and the cross-trend index , and define the following online update loss function :
[0034] where: represents the prediction output by the current model for the input under the parameter (such as the feature vector of a paper or a scholar node); is a custom high-order norm; is the exponent to amplify the prediction deviation; represents the cross-trend measure corresponding to the input , which can be mapped by or hot-spot markers. When is more associated with the hot topic, takes a larger value; , is the sample preference reinforcement index, used to intensify or weaken the preference for high samples; Through methods such as gradient descent or variational inference, the online update loss function is iteratively solved to update the model parameters, and the finally obtained online update model carries the adaptability to the latest academic hotspots;
[0035] where, is the step size coefficient of online learning, which determines the amplitude of each incremental update; When used, the hot topic list and the cross-trend index are used to assign higher training weights to emerging topics, so that the model can quickly learn the latest research trends under the condition of limited incremental data and quickly adapt to new trends; The weight assignment strategy is different from the traditional uniform loss. This formula dynamically adjusts the importance of each sample according to the cross-trend index and can better capture the rapidly emerging research directions.
[0036] Step 302. Transfer learning and environment adaptation If it is necessary to deploy in different universities or regions, there are often differences in subject layout and data distribution, and the hot topics and cross-trends are not exactly the same in different environments. Here, by analyzing the node distribution and external environment labels in the incremental knowledge graph , the source domain (such as the environment of a research-intensive university) and the target domain (such as the environment of a teaching-oriented university) are distinguished; To characterize the distribution difference between the source environment and the target environment, a transfer loss function is defined :
[0037] where: is the parameter vector or matrix of the evaluation model (possibly from online learning or initial training), which acts on the input to produce an output representation or prediction result; , are the datasets of the source environment (such as a research-intensive university) and the target environment (such as a teaching-oriented university), respectively; is the node or sample after being mapped by the model parameter to obtain the embedded vector or prediction output; is the transformation matrix or operator for cross-domain alignment, which performs a weighted transformation on and can be learned simultaneously during transfer optimization or allowed to vary piecewise over time; is a high-order or weighted vector norm (consistent with the previous one), which is used to amplify or weighted distance calculation for . The common one is norm ( is more sensitive when is the power exponent, taking a value greater than 1; is the weighted coefficient of the cross-domain sample pair and can be assigned according to external priors (such as the similarity between the two domains, cross-trend indicators, etc.). , if this item is not used, it can be set to ; If the transfer loss function is greater than expected, it means that more aggressive transfer is needed to adapt to the target environment. Based on the existing online updated model in the source environment, the incremental data in the target environment is subjected to transfer optimization, and a comprehensive objective function is defined:
[0038] where: is similar to the online learning loss form in step 301, except that here it is for the incremental data in the target environment; is the domain difference measurement term defined above, for specific reference, see above; is the balance coefficient, which controls the balance degree between the two during the transfer learning process; By minimizing , on the one hand, it ensures the prediction accuracy of the target domain data, and on the other hand, it reduces the conflict with the feature distribution of the source domain model, thereby realizing environment-adaptive transfer learning; After gradient solving or other optimization algorithms are used to solve, an optimized model is obtained , which has high adaptability to the emerging interdisciplinary trends in the target environment.
[0039] When in use, in different types of universities or research institutions, instead of blindly following the unified model, the distribution difference is measured through the domain difference measurement item , and targeted migration is carried out after measuring the distribution difference, which significantly improves the credibility of the evaluation results and realizes cross-domain adaptation; on the basis, the general features learned in the source environment are retained, and the existing accumulation will not be completely abandoned due to domain change, so that the experience can be brought to the new field, and the source model experience can be retained; through balancing the differences between the target domain and the source domain, the cross-domain generality or local accuracy can be flexibly selected to adapt to the needs of various actual scientific research environments.
[0040] Integrate the cross-trend index with the hot topic list into the update mechanism of online learning and transfer learning organically, so as to realize the adaptive optimization of the model in the situation where new data emerges continuously and there are multiple environmental differences; Step 301 focuses on quickly adapting to sudden academic hotspots in the same environment, and Step 302 further enables the model to adapt to the local characteristics of different universities or research institutions; through the completion of this step, the optimized model can not only continuously maintain high-precision prediction in the dynamically evolving scientific research environment, but also provide real-time and reliable basis for the fourth step of scenario simulation and strategy verification, forming a closed loop from data collection to trend recognition and then to model iteration, and truly achieving the forward-looking and long-term evaluation of the scientific research intersection potential.
[0041] Step Four: When the optimized model is obtained by solving and the management terminal submits the combination of external environment parameters and scientific research orientation variables to obtain the scenario , the scenario simulation framework immediately calls the scenario potential function , and outputs each potential result after performing time series iteration and high-order norm measure on the discipline evolution; The fourth step includes the following contents: Step 401: Definition and parameter configuration of multi-scenario environment According to the optimized model and the incremental knowledge graph obtained in the previous step, define the external environment parameters and scientific research orientation variables to be simulated here, including: External environmental parameters : It can cover the intensity of government funding, changes in industry demand, impact of public emergencies, etc.; Research Orientation Variables :Such as university funding allocation strategies, key discipline support plans, and interdisciplinary joint laboratory construction; For different combinations Assign a unique identifier and external environment parameters and research orientation variables Associate with the subject nodes or cooperation networks in the incremental knowledge graph. For example, weighting coefficients can be set on graph feature dimensions such as cooperation frequency, citation speed, or the rate of introduction of new disciplines to reflect the degree of intervention brought about by a specific environment or orientation.
[0042] Select simulation duration or key time period , such as the next 3 or 5 years, and in the simulation time The internal segment sets the environment or guidance change nodes to form a scene function that evolves over time , used to dynamically load scene configuration; where: Scene Function In the simulation area is defined in piecewise constant vector form:
[0043] in: It is a time segmentation point selected according to the guidance release time, funding disbursement cycle or major event nodes, which meets the ; It is The environmental orientation vector of a segment is composed of the external environmental parameters (such as the annual research funding ratio, the industry demand intensity, etc.) and the orientation variables (such as the discipline support weight and the cooperation incentive degree) within the segment; is the interval indicator function, when When it is 1, otherwise it is 0; In practical applications, It can correspond to key moments such as the launch date of major national scientific research projects, the reallocation date of university budgets, or the outbreak date of public health emergencies; and The data will be normalized or directly assigned a value based on the proportion of indicators such as the current document output, the number of projects approved or the amount of special funding.
[0044] in this way, This can accurately reflect the multi-stage guidance and environmental changes that evolve over time, and is used to call the optimized model in sub-step 402 Provide clear and quantifiable inputs during scenario simulation; Based on the above information, output a set of scenario definitions , each scenario corresponds to a set of environment-orientation combinations and the details of their evolution over time.
[0045] During use, by assigning unique identifiers to different orientation and environment parameters and performing mapping, subsequent simulations can flexibly call various scenario definitions to achieve rapid switching between multiple sets of assumptions under the same model framework; with the help of the optimized model and the atlas and trend information in the first and second steps, the scenario definitions can accurately grasp the key intervention points of real scientific research activities.
[0046] Step 402, Multi-scenario simulation and calculation of interdisciplinary potential After obtaining the set of scenario definitions , to uniformly analyze each scenario , introduce a simulation driver function to call the optimized model to evaluate the potential of the discipline nodes or cooperation networks at time . Among them, the simulation driver function refers to the mapping function that, after inputting the scenario and time into the optimized model , batch-calculates the potential of all nodes, and it can be specifically defined as:
[0047] where; is the concatenated feature vector of all nodes in the atlas at time ; is the scenario vector defined in sub-step 401 (the concatenation of environmental parameter and orientation variable ); projects the scenario vector into the node feature space; represents vector concatenation; and are the weights and biases of the first layer of the network, is the ReLU activation; and are the weights and biases of the second layer of the network, is a smooth non-linear mapping such as Softplus or Sigmoid; The output of the th component is the node The cross-potential prediction value at this scenario and moment; This structure can not only integrate the time-point features and scenario intervention factors, but also ensure the capture of complex cross-influences through multi-layer non-linear mapping, providing a refined and differentiable driving signal for the time integration and strategy comparison in sub-step 402.
[0048] Inside the simulation driving function, the external environment parameters and the scientific research orientation variables will be comprehensively used to perform weighted correction on relevant features, and then the updated node embedding feature vector representation will be sent into the optimized model , so as to output a new cross-potential prediction value; For the time series evolution, piecewise iteration or interpolation methods can be adopted to let the simulation driving function complete a potential calculation in combination with the scenario configuration in each sub-period, and then accumulate it to the next period until the entire simulation duration is covered ; To quantitatively evaluate the disciplinary potential trends under different scenarios , a scenario potential function is defined :
[0049] Where: are the parameters of the optimized model, called by the simulation driving function to calculate or predict the potential of each disciplinary node at any moment under the scenario : is the environmental one-oriented combined scenario; is the simulation driving function, which outputs the potential vector at the moment under the scenario ; where represents the potential value of the node (such as a certain discipline or cooperation group); is the partial derivative of the potential value of the node with respect to time (or can be approximated by difference under discrete time conditions), representing the change rate of the node potential as time progresses; is the set of nodes determined to be the key focus of disciplinary cross at the scenario and the moment ; is the high-order or mixed vector norm, which is of type ( ) or other mixed-type norms; is the amplification factor, or larger; is a time-weighted function, to is the simulation interval; where, ; Meaning: with the center of the simulation interval as the highest weight point, symmetrically decaying forward and backward; Parameters: , determines the position of the center of gravity (for example is the midpoint of the interval); , determines the width (the larger, the slower the decay); Effect: When it is desired to focus on evaluating the mid-term of the simulation or a critical moment, it can smoothly focus on that period, giving lower weights to moments farther from the center of gravity; Scenario potential function The larger it is, the more likely it is to achieve a breakthrough in the potential for interdisciplinary or the output of achievements in this scenario. It can also reverse-evaluate the cost-benefit ratio of the implementation of the guiding orientation according to requirements; for the scenario definition set in all , calculate the scenario potential function respectively, and then sort or classify and summarize to obtain the multi-scenario potential comparison for strategy screening and risk assessment.
[0050] When used, through the scenario definition set and the simulation drive function , it is possible to test the impacts of multiple combinations of environments and guiding orientations under a unified process, enabling the management side to intuitively compare different decision-making schemes, integrate the potential assessment over time, observe the differences between short-term and long-term effects, avoid only focusing on extreme values at a certain time while ignoring cumulative gains or delayed effects, use the optimized model optimized by the third-step online and transfer learning model for potential prediction, and then weight the graph features in combination with scenario intervention parameters to ensure that the simulation results have both accuracy and executability.
[0051] Step 403, Strategy Screening and Iterative Feedback According to the multi-scenario potential comparison , combined with the actual management requirements, define the strategy benefit function , such as combining the multi-scenario potential comparison with factors such as input resources and the difficulty of guiding implementation, calculate the benefit-risk balance value for each scenario for sorting or clustering; If a certain scenario If the potential is high but the implementation difficulty and risk are equally high, the management side can make a cautious choice in the final decision-making. If some other scenarios achieve a better balance between potential and risk, they can be given priority to enter the actual deployment stage, so as to realize the selection of scenarios ; In the scenario potential function When significant differences occur or obvious deviations from the expectations appear, the deviation information and external feedback in the simulation process can be collected into a new batch of incremental data and sent back to the online and transfer learning model optimization stage in the third step for model correction or weighted strengthening. This iterative feedback mechanism ensures that when the gap between the simulation and reality is large, the model can update its parameters in a timely manner , and be closer to the actual situation in subsequent simulations or real deployments. Finally, step 403 outputs a strategy plan and a risk list, including a priority strategy combination (such as increasing investment in a certain cross-field at a specific time point) and response plans under different environmental assumptions; When in use, the deviation of the simulation results and real data are fed back to the optimized model in the third step through the retraining mechanism , continuously improving the fitness of the optimized model and avoiding becoming quickly outdated as the environment changes after a one-time evaluation. By comprehensively considering scenario potential, resource consumption, and execution risk, the management side can allocate limited resources more rationally, avoid blindly pursuing a certain high-potential field while ignoring the implementation difficulty, and achieve multi-objective balance.
[0052] Integrate the optimized model with the incrementally accumulated knowledge graph and cross-trend indicators in the early stage, build a customizable multi-scenario simulation framework, quantify the impacts of different orientations and external environments, and output targeted strategies and risk assessments on this basis.
[0053] Combined with the online and transfer learning mechanism in the third step, a complete closed-loop of data collection - trend recognition - model optimization - strategy simulation is finally formed, ensuring that the assessment of the potential of scientific research intersections is both forward-looking and keeping pace with the times, providing strong technical support for universities and research institutions to make accurate decisions in a dynamic environment.
[0054] However, with the increasing diversification of the forms of scientific research activities and data sources, a single text or citation network is no longer sufficient to fully capture the interaction and knowledge diffusion laws of some emerging disciplines. Therefore, more modalities of data (such as academic conference videos, experimental process data, personal social media interactions, etc.) are introduced and deeply integrated with the existing incremental knowledge graph to further enrich and improve the characterization of cross-research activities.
[0055] Step Five: When a significant increment in multi-modal sources is detected based on the graph text, for the generated multi-modal original dataset Execute modal marking and semantic vector extraction, and then pass through an aggregation function Align cross-modal features and output a multi-modal fusion library Finally, incrementally integrate it into the multi-modal knowledge graph To strengthen the stereoscopic recognition of emerging cross-disciplinary fields and the adaptability of online models; Step five includes the following content: Step 501, Multi-modal data acquisition and preprocessing Based on the text, patents, citation networks, etc. used in the previous four steps, introduce multi-modal information such as videos, audios, images, and sensor logs. Incorporate these new sources into the data acquisition pipeline uniformly and generate a multi-modal raw dataset using their respective corresponding acquisition strategies (such as regularly capturing live replays of academic conferences and obtaining output logs of scientific research equipment). ; For the multi-modal raw dataset Conduct preliminary modal discrimination and extract core metadata (such as video duration, audio transcription text, image scene description, etc.). Then, perform semantic annotation on this core metadata information to generate multi-modal marked data ; This step will be combined with text mining to discover disciplinary keywords, research institutions, scholar names, etc. mentioned in video or audio content. If the same event (such as a certain scientific research forum) appears in multiple modalities such as conference videos, social media posts, and paper citations, it is necessary to align through multiple factors such as timestamps, keywords, and geographical locations and merge them into the same entity or relationship node to avoid repeated insertion in the incremental knowledge graph.
[0056] At this time, refer to the concept drift detection mechanism in sub-step 102 to mark the sudden associations that may be brought by new modalities and output them to the temporary multi-modal structure ; Using non-text modalities such as videos, audios, and images as supplements to achieve a more comprehensive capture of academic activities, thereby expanding the data coverage. Through multi-angle matching of timestamps, metadata, etc., avoid knowledge graph structure conflicts caused by repeated mapping of different modalities.
[0057] Step 502, Multi-modal embedding representation and semantic fusion Use deep learning models (such as CNN, Transformer, or hybrid neural structures) to extract feature vectors for data segments of each modality (video frames, audio segments, image regions, etc.) to form a temporary multi-modal vector set , if it involves text transcription or OCR (graphic and text recognition) content, the semantic labels can be fused with visual / audio features by combining the entity information of the aforementioned text mining; Introduce cross-modal alignment transformation for features of different modalities Perform mapping and aggregate within a unified representation space, and define the following multi-modal aggregation function :
[0058] where: is the vector representation of the th multi-modal segment at time ; is the modal weight factor used to strengthen or suppress the importance of certain modalities; is a linear or non-linear mapping used to complete alignment in the unified space; is the higher-order or mixed norm defined previously; is the amplification exponent; The aggregation function outputs the multi-modal semantic vector , representing the comprehensive features of a certain entity (such as a scholar or a research team) in each modality, and is used for the next step of knowledge graph merging; after completion of alignment and aggregation, the multi-modal semantic vector is cross-checked with the keyword tags and hot candidate lists generated by text mining.
[0059] If abnormal noise is found (for example, the image recognition result strongly conflicts with the text entity), enable a concept drift or inconsistency detection mechanism similar to that in step 102 for correction, and the finally retained multi-modal features will be uniformly stored in the multi-modal fusion library ; When in use, extract interdisciplinary traces from multiple aspects of video / audio / text to further enhance the sensitivity to emerging scientific research fields; it can adapt to the usage requirements of multi-modal information in different universities and different fields, coordinate with the keyword analysis in the second step, the online learning in the third step, etc., and add the new multi-modal semantic vector to the input feature space of the subsequent evaluation model, which helps the optimized model to be further iterated in subsequent re-learning.
[0060] Step 503, Multi-modal incremental integration and dynamic update Utilize the multi-modal fusion library to create / update the corresponding nodes or relationships in the incremental knowledge graph for each newly emerged or significantly changed entity (researcher, experimental project, video topic, etc.). The update rules are the same as the incremental merging strategies adopted in the first and second steps, but the attribute fields need to be expanded (such as adding visual description speech tags, etc.) to record multi-modal information.
[0061] Record the updated multi-modal knowledge graph as , and introduce the online learning and transfer learning in the third step and the scenario simulation and strategy verification mechanism in the fourth step for rapid verification: If the multi-modal information indicates a certain hot topic with a sharp increase in the frequency of occurrence in the video or experimental log, it can trigger the fine-tuning of the optimized model ; if the new modality shows that the collaborative research methods in a certain field are more intensive, the corresponding potential value or the proportion of directed investment can also be increased in the simulation session; when the multi-modal update causes changes in the interdisciplinary trend or the hot list, these changes and potential deviation information need to be fed back to step 201, 202 or the third online learning module for regular or real-time re-training, and a new strategy simulation can also be triggered when necessary, so that the management side can understand the differences brought by the multi-modal information to the evaluation results and resource allocation suggestions.
[0062] When in use, it can keep the incremental knowledge graph timely collect new modality entities, which can not only merge text data, but also be compatible with new interdisciplinary signals generated by information such as video and audio; after multi-modal fusion, some potential breakthrough points that are not easily detected in traditional text data (such as abnormal laboratory instrument data and the popularity of group discussions at academic conferences) can be quickly converted into monitorable nodes to strengthen the forward-looking perception of hot fields; it realizes the comprehensive collection and deep fusion of non-text information such as images, audio, video, and sensor logs, and feeds the newly obtained multi-modal features back to the optimized model and the scenario simulation framework to further improve the perception and decision-making support capabilities of this system for the rapidly iterating, cross-modal scientific research ecosystem, and provide more comprehensive and three-dimensional interdisciplinary insights for the scientific research management side of universities.
[0063] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but this implementation should not be considered to exceed the scope of this application.
[0064] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0065] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.
[0066] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0067] As mentioned above, the above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. An interdisciplinary scientific research potential evaluation method based on a dynamic multi-modal knowledge graph, characterized in that: including If the newly crawled academic data volume exceeds the threshold, text mining is used to extract entities and relationships. After identifying emerging terms through concept drift detection, they are incrementally merged into the incremental knowledge graph, and then the ontology level is expanded to accommodate new concepts; After reaching a new time slice, calculate the multi-time window differential vector, aggregate it to highlight the theme growth or convergence trend, generate cross-trend indicators and a list of hot topics, and mark sudden rises and falls; If the cross-trend indicators and the list of hot topics show significant fluctuations within the domain, perform online parameter updates to optimize the model parameters, and use the transfer loss function to adaptively fine-tune the model parameters in different regions; When the optimized model is obtained and the management end submits a scenario composed of external environmental parameters and scientific research orientation variables, call the scenario potential function to evaluate the potential trend of resource allocation and interdisciplinary collaboration, output comparable indicators, and when the actual and predicted differences are greater than expected, feedback a deviation signal to the online learning module; Perform modal marking and semantic vector extraction on the generated multi-modal original dataset, align the cross-modal features by the aggregation function and output the multi-modal fusion library, incrementally integrate it into the multi-modal knowledge graph, and trigger re-fine-tuning of the optimized model when anomalies occur.
2. The interdisciplinary scientific research potential evaluation method according to claim 1, characterized in that: Establish a timed or real-time crawling strategy for multi-source data, set the crawling frequency and filtering conditions, and then obtain the staged original dataset; use text mining algorithms to perform entity annotation on the common structures in the academic fields in the staged original dataset, and preliminarily merge duplicate entities or name variants to obtain an entity set, and form a preliminary relationship set after extracting the association information between the documents.
3. The interdisciplinary scientific research potential evaluation method according to claim 2, characterized in that: Incrementally merge the entity set and the relationship set into the incremental knowledge graph. If overlap with existing nodes or attribute conflicts are detected, a merging strategy based on a confidence threshold is adopted for correction; Perform concept drift calculation for nodes and edges in the incremental knowledge graph; If the obtained concept drift measure exceeds the preset drift threshold, use it as a new concept node, trigger the corresponding ontology expansion or hot topic marking process, incorporate the uncollected interdisciplinary fields into the semantic structure of the graph, and mark them as hot topic candidates, and output a list of hot topic candidates.
4. The interdisciplinary scientific research potential evaluation method according to claim 1, characterized in that: Add marks to the list of hot topic candidates so that it includes an additional cross-potential dimension when generating dynamic feature vectors; Based on the incremental knowledge graph, aggregate and generate serialized graph slices, extract key information from each graph slice, generate feature vectors of nodes through embedding learning, and extract the differential or high-order change rate of the feature vectors of adjacent graph slices to form a time-series differential vector.
5. The interdisciplinary scientific research potential evaluation method according to claim 4, characterized in that: For each topic node, calculate the aggregate function that measures the overall evolution trend across time periods; When the trend aggregation function exceeds the preset aggregation threshold, determine the corresponding topic as a node with significant cross-potential and mark it as a cross-trend indicator. Summarize it into the global trend information at the entire graph dimension, combine the time-series differential vector and the cross-trend indicator, screen the detected new concept nodes, and obtain a list of hot topics.
6. The interdisciplinary scientific research potential evaluation method according to claim 5, characterized in that: After obtaining new academic information and combining it with the list of hot topics, select nodes and relationships with high relevance to the new hot topics in the incremental knowledge graph and construct an incremental training set; Combined with the incremental training set and cross-trend indicators, an online update loss function is constructed, and the online update loss function is iteratively solved. After updating the model parameters, an online update model is obtained.
7. The interdisciplinary scientific research potential evaluation method according to claim 6, characterized in that: By analyzing the node distribution and external environment labels in the incremental knowledge graph, after distinguishing the source domain and the target domain, a transfer loss function is defined to characterize the distribution difference between the source environment and the target environment; If the transfer loss function is greater than the expected value, based on the existing online update model in the source environment, the incremental data in the target environment is transferred and optimized, and the gradient of the defined comprehensive objective function is solved to obtain an optimized model.
8. The interdisciplinary scientific research potential evaluation method according to claim 7, characterized in that: After collecting the external environment parameters and scientific research orientation variables, they are associated with the theme nodes or cooperation networks in the incremental knowledge graph; Change nodes of the environment or orientation are set in segments within the simulation duration to form a scenario function that evolves over time, which is used to dynamically load scenario configurations, and finally a scenario definition set containing several scenarios is output.
9. The interdisciplinary scientific research potential evaluation method according to claim 8, characterized in that: The updated node embedding feature vector representation is fed into the optimized model to output a new cross-potential prediction value; The simulation-driven function calculates the potential in each sub-period combined with the scenario configuration, and then accumulates it to the next period until the entire simulation duration is covered; the scenario potential is calculated for all scenarios in the scenario definition set respectively, and then sorted or classified and summarized to obtain a comparison of multi-scenario potentials.
10. The interdisciplinary scientific research potential evaluation method according to claim 9, characterized in that: Combined with the actual management requirements, a policy benefit function is defined to select scenarios. When there are significant differences in scenario potential or obvious deviations from the expected value, the deviation information and external feedback during the simulation process are collected into a new batch of incremental data for model correction or weighted reinforcement, and a policy plan and a risk list are output, including the priority policy combination and response plans under different environmental assumptions.
11. The interdisciplinary scientific research potential evaluation method according to claim 10, characterized in that: After introducing multi-modal information into the data acquisition pipeline, a multi-modal raw data set is generated using their respective acquisition strategies; the multi-modal raw data set is preliminarily modality-distinguished and core metadata is extracted, and then semantic annotation is performed to generate multi-modal labeled data. If the same event appears in multiple modalities, it is aligned through multiple factors and merged into the same entity or relationship node.
12. The interdisciplinary scientific research potential evaluation method according to claim 11, characterized in that: Feature vectors are extracted from the data segments of each modality using a deep learning model to form a temporary multi-modal vector set; A cross-modal alignment transformation is introduced to map the features of different modalities and aggregate them in a unified representation space to output multi-modal semantic vectors; the multi-modal semantic vectors are cross-checked with the keyword labels and hot candidate lists generated by text mining. If abnormal noise is found, a concept drift or inconsistency detection mechanism is enabled for correction.
13. The interdisciplinary scientific research potential evaluation method according to claim 12, characterized in that: Corresponding nodes or relationships in the incremental knowledge graph are created or updated for each newly emerged or significantly changed entity, and an online learning and transfer learning and scenario simulation and policy verification mechanism is introduced for rapid verification. If the multi-modal information indicates that the frequency of a certain hot topic surges in videos or experimental logs, fine-tuning of the optimized model is triggered; If the collaborative research method in a certain field is more intensive, the corresponding potential value or investment ratio is increased in the simulation link. When the multi-modal update causes changes to the interdisciplinary trend or hot list, the corresponding changes and potential deviation information are fed back again.
Citation Information
Patent Citations
Power transmission and transformation project knowledge graph construction method based on machine learning
CN114936641A
Intelligent doorbell visitor identification method and system based on deep learning
CN118761034A
Cited By
Adaptive AI control method for fusing heterogeneous data based on multi-modal perception
CN120469255A
An adaptive AI control method based on multimodal perception and fusion of heterogeneous data
CN120469255B
Aero-engine systematic multi-dimensional digital characterization method and system
CN120509320A
Examination proposition person behavior model construction method and system based on deep learning
CN120822160A
Basic scientific research institution hotspot analysis terminal based on edge calculation and lightweight AI and implementation method
CN121167356A
