Method and apparatus for classifying scientific research projects based on personalized machine learning

Abstract

The application relates to the field of scientific research project management, and discloses a scientific research project classification method and device based on individualized machine learning, which comprises the following steps: determining a scientific research interest portrait and a preference weight of a user; determining labeled scientific research projects corresponding to a plurality of scientific research projects to be classified according to the scientific research interest portrait, the preference weight, real-time field hotspot data and scientific research project attribute data of the plurality of scientific research projects to be classified; determining matchable classification labels corresponding to the plurality of scientific research projects to be classified; determining matching deviation values corresponding to the plurality of scientific research projects to be classified; and performing demand verification optimization on current scientific research demands of the user and the matching deviation values corresponding to the plurality of scientific research projects to be classified to obtain a demand adaptation classification set in which the plurality of scientific research projects to be classified adapt to the current scientific research demands. According to the application, scientific research project classification can accurately match the research direction, past preferences and current scientific research stage of a user.

Description

Technical Field

[0001] This application relates to the field of scientific research project management technology, and more specifically, to a method and apparatus for classifying scientific research projects based on personalized machine learning. Background Technology

[0002] Existing methods for classifying research projects largely rely on fixed research project systems. These systems typically use general dimensions such as discipline, research field, and technical direction as the basis for division, with pre-set unified classification standards and hierarchical structures. In practical applications, existing classification methods mainly extract core attribute information of research projects and compare it one by one with pre-set classification standards to complete project classification. This method is widely used in scenarios such as research management archiving, preliminary screening of project searches, and overall coordination of research resources. Existing classification methods have fixed logic and standardized operating procedures, enabling large-scale classification and organization of research projects and meeting the basic needs of general research management for project classification.

[0003] Existing research project classifications struggle to accurately match users' research interests, past preferences, and current research stages. Because fixed classification systems use uniform standards, they fail to incorporate personalized insights based on users' historical research data, resulting in classifications that are highly general but lack specificity. For users at different research stages and with varying research preferences, existing classifications often include numerous projects with low relevance to their needs, requiring users to invest additional time and effort in secondary filtering, significantly increasing research costs. Summary of the Invention

[0004] The purpose of this application is to provide a scientific research project classification method and apparatus based on personalized machine learning, which solves the technical problem that it is difficult to accurately match the user's research direction, past preferences and current research stage when classifying scientific research projects, and achieves the technical effect that the scientific research project classification can accurately match the user's research direction, past preferences and current research stage.

[0005] In a first aspect, embodiments of this application provide a research project classification method based on personalized machine learning. The method includes: acquiring a user's historical research data and research stage tags; extracting features from the historical research data and research stage tags to determine the user's research interest profile and preference weights; wherein the historical research data includes followed projects, classification habits, and research results; acquiring real-time domain hotspot data and research project attribute data for multiple research projects to be classified; using a domain heat labeling unit, determining the labeled research projects corresponding to multiple research projects to be classified based on the research interest profile, preference weights, real-time domain hotspot data, and research project attribute data for multiple research projects to be classified; and classifying multiple research projects using a preset project classification system. The system classifies and evaluates labeled research projects, resulting in multiple matching category tags for these projects. Each labeled research project has real-time tags and popularity values. Multi-dimensional matching and filtering are performed on the user's research interest profile and the matching category tags for the multiple projects to be classified, yielding matching deviation values ​​for each project. The system then identifies the user's current research needs. Finally, it performs requirement verification and optimization on the user's current research needs and the matching deviation values ​​for the multiple projects to be classified, resulting in a set of requirement-adaptive categories that fit the current research needs. This set includes multiple requirement-adaptive categories that fit the current research needs, including project applications and literature reviews.

[0006] In one possible implementation, feature extraction is performed on historical research data and research stage tags to determine the user's research interest profile and preference weights. This includes: obtaining the user's research field, historical classification records, and published paper topic data from the platform's user research database based on the user's unique identifier; performing One-Hot encoding and TF-IDF weight statistics on the user's research field, historical classification records, and published paper topic data to determine personalized feature vectors and a domain preference weight table; using the personalized feature vectors as the research interest profile and the domain preference weight table as the preference weights; performing multi-dimensional matching and filtering on the user's research interest profile and the matching classification tags corresponding to multiple research projects to be classified to obtain matching deviation values ​​for multiple research projects to be classified, including: embedding the personalized feature vectors and the domain preference weight table into a unified feature space; determining the domain preference weight vectors corresponding to the matching classification tags corresponding to the research projects to be classified within the unified feature space; and determining the sum of the distances between the personalized feature vectors and the domain preference weight vectors, and the distances between the vectors corresponding to the domain preference weight table and the domain preference weight vectors within the unified feature space, as the matching deviation values ​​for multiple research projects to be classified.

[0007] In another possible implementation, the method further includes: acquiring real-time research interaction data of users; determining multiple peer collaboration features based on personalized feature vectors, domain preference weight tables, and real-time research interaction data using a rule-based neighbor routing unit; wherein, real-time research interaction data includes browsing projects, collecting literature, and peer collaboration records; concatenating multiple peer collaboration features and personalized feature vectors to obtain a fusion behavior feature vector; acquiring incremental research interaction data of users; iteratively updating the fusion behavior feature vector and incremental research interaction data using sliding window statistics to obtain an updated fusion behavior feature vector; wherein, the peer collaboration features and personalized feature vectors are concatenated with a 7:3 weight; mapping the updated fusion behavior feature vector and the research project attribute data of the research projects to be classified to the same feature space and standardizing them; determining the user-project matching values ​​corresponding to multiple updated fusion behavior feature vectors and the research projects to be classified through cosine similarity matching; determining multiple target user-project matching values ​​greater than or equal to the preset user-project matching values, and using the target updated fusion behavior feature vectors corresponding to the multiple target user-project matching values ​​as research interest profiles.

[0008] In another possible implementation, a rule-based neighbor routing unit determines multiple peer collaboration features based on personalized feature vectors, a domain preference weight table, and real-time research interaction data. This includes: determining the user's core research domain through the domain preference weight table; identifying multiple target real-time research interaction data sets that match the core research domain, meet collaboration frequency requirements, and have valid collaboration types from the real-time research interaction data using preset rules; determining collaboration weights based on the collaboration frequencies corresponding to the multiple target real-time research interaction data sets; and multiplying the domain preference weights in the domain preference weight table by the collaboration weights to create a modified domain preference weight table, which is then used as the multiple peer collaboration features.

[0009] In another possible implementation, determining multiple target user-project matching values ​​that are greater than or equal to a preset user-project matching value includes: obtaining users' historical classification feedback data based on their unique user identifiers from the platform's user research database; determining a project classification screening threshold based on the historical classification feedback data; determining a research stage adjustment coefficient based on research stage tags; determining the product of the project classification screening threshold and the research stage adjustment coefficient as the stage project classification screening threshold; and using the stage project classification screening threshold as the preset user-project matching value to determine multiple target user-project matching values ​​that are greater than or equal to the preset user-project matching value.

[0010] In another possible implementation, the domain popularity labeling unit is trained as follows: The sample research interest profile, sample preference weights, real-time domain hot data, research project attribute data of multiple research projects to be classified, real-time labels of the labeled research projects, and sample popularity values ​​are mapped to a unified vector space; the domain popularity labeling unit is trained using the sample research interest profile, sample preference weights, real-time domain hot data, research project attribute data of multiple research projects to be classified, real-time labels of the labeled research projects, and sample popularity values; wherein, the domain popularity labeling unit includes an input layer, a feature alignment layer, a two-branch matching layer, and a fusion calculation layer arranged sequentially, and the two-branch matching layer includes user interest matching. The branch layer and the domain hotspot association branch layer; the user interest matching branch layer is used to determine the basic matching degree between the research project attribute vector corresponding to the research project attribute data and the research interest vector corresponding to the research interest profile, and is used to determine the product of the basic matching degree and the preference weight vector corresponding to the preference weight, as the weighted interest matching degree; the domain hotspot association branch layer is used to determine the basic correlation degree between the research project attribute vector and the real-time domain hotspot vector corresponding to the real-time domain hotspot data, and inputs the basic correlation degree and the timeliness coefficient into the time decay weighting module, and the time decay weighting module outputs the corrected hotspot correlation degree; the weighted interest matching degree, the corrected hotspot correlation degree, the real-time labels of multiple sample labeled research projects and the vectors corresponding to the sample popularity values ​​are input into the fusion calculation layer.

[0011] In another possible implementation, the user's current research needs and the matching deviation values ​​corresponding to multiple research projects to be classified are optimized through demand verification. This yields a demand adaptation category set for multiple research projects to be classified that fits the current research needs. The demand adaptation category set includes multiple demand adaptation categories that fit the current research needs, including: obtaining the feasibility weight, the first domain popularity weight, and the first matching deviation weight; when the current research need is a project application, obtaining the feasibility indicator and the first domain popularity corresponding to the current research need; determining the sum of the product of the feasibility indicator and the feasibility weight, and the product of the first domain popularity and the first domain popularity weight, as the feasibility evaluation indicator; determining the product of the matching deviation value and the first matching deviation weight, as the matching deviation evaluation indicator; determining the difference between the feasibility evaluation indicator and the matching deviation evaluation indicator, as the comprehensive evaluation indicator for multiple research projects to be classified; and identifying multiple research projects to be classified whose comprehensive evaluation indicator is greater than or equal to the preset comprehensive evaluation indicator, thus obtaining a demand adaptation category set for multiple research projects to be classified that fits the current research needs.

[0012] In another possible implementation, the user's current research needs and the matching deviation values ​​corresponding to multiple research projects to be classified are optimized through demand verification. This yields a demand adaptation classification set that adapts multiple research projects to the current research needs. The demand adaptation classification set includes multiple demand adaptation classifications that adapt to the current research needs, including: obtaining academic relevance weights, secondary field popularity weights, and secondary matching deviation weights; when the current research need is literature review, obtaining the academic relevance indicators and secondary field popularity corresponding to the current research need; determining the sum of the product of the academic relevance indicators and academic relevance weights, and the product of the secondary field popularity and secondary field popularity weights, as the academic relevance evaluation indicator; determining the product of the matching deviation value and the secondary matching deviation weights, as the matching deviation evaluation indicator; determining the difference between the academic relevance evaluation indicator and the matching deviation evaluation indicator, as the comprehensive evaluation indicator for multiple research projects to be classified; and identifying multiple research projects whose comprehensive evaluation indicator is greater than or equal to the preset comprehensive evaluation indicator, thus obtaining a demand adaptation classification set that adapts multiple research projects to the current research needs.

[0013] In another possible implementation, the feasibility weight is 0.3, the popularity weight of the first field is 0.2, and the first matching deviation weight is 0.5; the academic relevance weight is 0.4, the popularity weight of the second field is 0.1, and the second matching deviation weight is 0.5.

[0014] Secondly, embodiments of this application provide a scientific research project classification device based on personalized machine learning, including units for implementing the above-described method.

[0015] The beneficial effects of the embodiments of this application compared with the prior art are:

[0016] This application provides a research project classification method based on personalized machine learning. The method includes: acquiring a user's research history data and research stage tags; extracting features from the research history data and research stage tags to determine the user's research interest profile and preference weights; acquiring real-time domain hotspot data and research project attribute data of multiple research projects to be classified; using a domain hotspot labeling unit, determining labeled research projects corresponding to multiple research projects to be classified based on the research interest profile, preference weights, real-time domain hotspot data, and research project attribute data of multiple research projects to be classified; classifying and evaluating multiple labeled research projects using a preset project classification system to obtain matching classification tags corresponding to multiple research projects to be classified; performing multi-dimensional matching and filtering on the user's research interest profile and the matching classification tags corresponding to multiple research projects to be classified to obtain matching deviation values ​​corresponding to multiple research projects to be classified; acquiring the user's current research needs; performing demand verification and optimization on the user's current research needs and the matching deviation values ​​corresponding to multiple research projects to be classified to obtain a demand adaptation classification set that adapts multiple research projects to the current research needs, the demand adaptation classification set including multiple demand adaptation categories that adapt to the current research needs. In this embodiment, the user's past research trajectory can be deeply explored, the user's research inclination can be accurately anchored, the deviation between the classification results and the user's interests can be effectively reduced, the personalization of the classification results can be greatly improved, and the classification can better match the user's fixed preferences formed by long-term scientific research accumulation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the first research project classification method based on personalized machine learning provided in this application embodiment;

[0019] Figure 2 A schematic diagram illustrating the workflow of the first research project classification method based on personalized machine learning provided in this application embodiment;

[0020] Figure 3 A flowchart illustrating the second method for classifying scientific research projects based on personalized machine learning, provided in this application embodiment;

[0021] Figure 4 A schematic diagram illustrating the workflow of the second research project classification method based on personalized machine learning provided in this application embodiment;

[0022] Figure 5 A flowchart illustrating the third method for classifying scientific research projects based on personalized machine learning, provided in this application embodiment;

[0023] Figure 6 A schematic diagram illustrating the workflow of the third research project classification method based on personalized machine learning provided in this application embodiment;

[0024] Figure 7 A flowchart illustrating the fourth method for classifying scientific research projects based on personalized machine learning, provided in this application embodiment;

[0025] Figure 8 A schematic diagram illustrating the workflow of the fourth research project classification method based on personalized machine learning provided in this application embodiment;

[0026] Figure 9 A flowchart illustrating the fifth research project classification method based on personalized machine learning provided in this application embodiment;

[0027] Figure 10 A schematic diagram illustrating the workflow of the fifth research project classification method based on personalized machine learning provided in this application embodiment;

[0028] Figure 11 A flowchart illustrating the sixth method for classifying scientific research projects based on personalized machine learning, provided in this application embodiment;

[0029] Figure 12 A schematic diagram illustrating the workflow of the sixth research project classification method based on personalized machine learning provided in this application embodiment;

[0030] Figure 13 A flowchart illustrating the seventh method for classifying scientific research projects based on personalized machine learning, provided in this application embodiment;

[0031] Figure 14 A flowchart illustrating the eighth scientific research project classification method based on personalized machine learning provided in this application embodiment;

[0032] Figure 15 A flowchart illustrating the ninth method for classifying scientific research projects based on personalized machine learning, provided in this application embodiment;

[0033] Figure 16 This is a schematic diagram of the logical structure of a scientific research project classification method system based on personalized machine learning, provided in an embodiment of this application. Detailed Implementation

[0034] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0035] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0036] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0037] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0038] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0039] Existing research project classification methods often result in a large number of projects that are not closely related to the needs of users at different research stages and with different research preferences. Users need to invest additional time and effort in secondary screening, which significantly increases the cost of research work.

[0040] Based on the above reasons, this application provides a research project classification method based on personalized machine learning. The method includes: acquiring a user's research history data and research stage tags; extracting features from the research history data and research stage tags to determine the user's research interest profile and preference weights; acquiring real-time domain hotspot data and research project attribute data of multiple research projects to be classified; using a domain hotspot labeling unit, determining labeled research projects corresponding to multiple research projects to be classified based on the research interest profile, preference weights, real-time domain hotspot data, and research project attribute data of multiple research projects to be classified; classifying and evaluating multiple labeled research projects using a preset project classification system to obtain matching classification tags corresponding to multiple research projects to be classified; performing multi-dimensional matching and filtering on the user's research interest profile and the matching classification tags corresponding to multiple research projects to be classified to obtain matching deviation values ​​corresponding to multiple research projects to be classified; acquiring the user's current research needs; performing demand verification and optimization on the user's current research needs and the matching deviation values ​​corresponding to multiple research projects to be classified to obtain a demand adaptation classification set that adapts multiple research projects to the current research needs, the demand adaptation classification set including multiple demand adaptation categories that adapt to the current research needs. In this embodiment, the user's past research trajectory can be deeply explored, the user's research inclination can be accurately anchored, the deviation between the classification results and the user's interests can be effectively reduced, the personalization of the classification results can be greatly improved, and the classification can better match the user's fixed preferences formed by long-term scientific research accumulation.

[0041] In some scenarios, a research project classification method based on personalized machine learning, as described in this application, can be applied to the classification and management of research projects. It can scientifically classify users' associated research projects, so as to recommend research projects to users in a scientific and reasonable way, thereby improving the efficiency of users in managing research projects.

[0042] The following section provides a detailed explanation of a research project classification method based on personalized machine learning, as provided in the embodiments of this application, using specific examples.

[0043] Figure 1 A flowchart illustrating the first research project classification method based on personalized machine learning provided in this application embodiment is shown below. Figure 1 As shown in the embodiment of this application, a scientific research project classification method based on personalized machine learning is provided. The method includes S110 to S130, and S110 to S130 are described in detail below.

[0044] S110. Obtain the user's research history data and research stage tags. Extract features from the research history data and research stage tags to determine the user's research interest profile and preference weights. The research history data includes followed projects, classification habits, and research achievements.

[0045] Figure 2 A schematic diagram illustrating the workflow of the first research project classification method based on personalized machine learning provided in this application embodiment is shown below. Figure 2 As shown, in this implementation, relevant data and stage markers generated by users in scientific research activities can be continuously collected. This information is the basic data source for constructing personalized classification criteria.

[0046] It should be noted that historical research data is a collection of a user's past research activities and achievements, while research stage tags are used to identify the current research progress node of the user.

[0047] For example, historical research data includes various academic projects tracked by the user in recent years and research categories compiled by the user themselves. The research stage tag can be marked as "Young Researcher Stage".

[0048] It should be noted that "projects followed" refers to various research project information actively tracked by users, "categorization habits" refers to users' self-defined categorization logic of research materials or projects, and "research results" refers to the research outputs completed by users.

[0049] For example, the projects of interest include national-level ongoing projects in a certain field. The classification convention is to divide the research content according to two major research dimensions, and the research results include published core journal papers and authorized patents.

[0050] In this implementation, feature extraction algorithms from machine learning can be used to decompose and extract information from the collected historical research data and research stage labels, construct a profile model that can accurately reflect the user's research inclinations, and assign corresponding preference weights to different research directions.

[0051] S120. Acquire real-time domain hotspot data and research project attribute data for multiple research projects to be classified. Using the domain heat labeling unit, based on research interest profiles, preference weights, real-time domain hotspot data, and research project attribute data for multiple research projects to be classified, determine the labeled research projects corresponding to the multiple research projects to be classified. Through a preset project classification system, classify and evaluate the multiple labeled research projects to obtain matching classification tags for the multiple research projects to be classified. The labeled research projects have real-time tags and heat values.

[0052] In this implementation method, popular and dynamic information in the current scientific research field can be collected through academic databases, industry research platforms and other channels. At the same time, various core attribute information of the scientific research projects to be classified can be collected to provide data support for subsequent classification and labeling.

[0053] It should be noted that real-time domain hotspot data is a collection of information on research directions and technological breakthroughs that are currently receiving widespread attention in the scientific research field.

[0054] For example, real-time domain hotspot data includes the latest research progress in a certain interdisciplinary field and the thematic content recently published in authoritative journals.

[0055] It should be noted that the research project attribute data refers to the various characteristic information of the research project itself, which is to be classified, covering the research direction, the platform on which it relies, and other content.

[0056] For example, the research project attribute data includes a research direction of green energy technology and the affiliated institution of a project to be classified as a key university.

[0057] In this implementation, a preset domain popularity labeling unit can be called to perform multi-dimensional fusion calculations on the user's scientific research interest profile, preference weights, real-time domain hot data, and attribute data of the project to be classified, so as to generate corresponding labeled scientific research projects for each scientific research project to be classified.

[0058] It should be noted that real-time tags are thematic identifiers assigned to labeled research projects based on user interests and current hot topics, while the popularity value is a quantitative indicator reflecting the degree of attention the project receives in the current research field and its relevance to user interests.

[0059] For example, the real-time tag for a research project can be "artificial intelligence, materials science", and the popularity value is the result of a comprehensive calculation of the field's attention and user preferences.

[0060] In this implementation method, a pre-built scientific research project classification system can be used to match and evaluate the labeled scientific research projects, and determine the appropriate classification label for each scientific research project to be classified.

[0061] It should be noted that the preset project classification system is a hierarchical classification framework built based on general classification standards in the scientific research field and users' potential classification habits. The classification evaluation process will combine the project's real-time tags, popularity value and category dimensions within the system for matching.

[0062] For example, the preset project classification system includes multiple primary research categories, and each primary category is further subdivided into multiple secondary categories. During the evaluation, the real-time tags of the research projects and the secondary categories are accurately matched to determine the corresponding matching classification tags.

[0063] S130. Perform multi-dimensional matching and filtering on the user's research interest profile and the matching category tags corresponding to multiple research projects to be classified, obtaining the matching deviation values ​​for multiple research projects to be classified. Obtain the user's current research needs. Perform demand verification and optimization on the user's current research needs and the matching deviation values ​​corresponding to multiple research projects to be classified, obtaining a demand adaptation category set that adapts multiple research projects to the current research needs. The demand adaptation category set includes multiple demand adaptation categories that adapt to the current research needs. Among them, the current research needs include project application and literature review.

[0064] In this implementation, the user's research interest profile and the matching classification tags of the projects to be classified can be compared and calculated from multiple dimensions such as research direction and technical field to obtain a quantitative value of the degree of deviation between each project to be classified and the user's interests.

[0065] It should be noted that the matching deviation value is an indicator that measures the degree of fit between the classification labels of the scientific research projects to be classified and the user's scientific research interest profile. The smaller the value, the higher the degree of fit.

[0066] For example, if the matching category label for a certain item to be classified is "quantum communication", and the user's interest profile has a low preference weight for this direction, then the matching deviation value of this item will be relatively high.

[0067] In this implementation, the user's current research task goals can be inferred from the user's actively input demand information or behavioral data. These demands will serve as the core guide for optimizing the classification results.

[0068] It should be noted that the current research needs refer to the user's specific research tasks in the near future, mainly covering two types: project application and literature review.

[0069] For example, the need for project application can be reflected in the user's need to match research project directions that meet the application requirements of a certain type of grant, and the need for literature review can be reflected in the user's need to collect cutting-edge project information under a certain research topic.

[0070] In this implementation, the matching deviation value of each project to be classified can be adjusted and verified in a targeted manner based on the user's current research needs, and finally a set of research project classifications that fit the user's current task can be selected.

[0071] It should be noted that the demand matching category set is a collection of multiple category results that are highly matched with the user's current research needs after demand verification and optimization. Each demand matching category corresponds to a type of research project that meets the demand.

[0072] For example, when a user's current research need is to apply for a research grant, the set of needs matching categories includes multiple categories that match different levels of funding projects.

[0073] This implementation method acquires users' historical research data and research stage tags. The historical research data includes projects followed, classification habits, and research results. Feature extraction is performed on this data and tags to determine the user's research interest profile and preference weights. Subsequently, the user's research interest profile and the matching classification tags corresponding to multiple research projects to be classified are matched and filtered in multiple dimensions to obtain the matching deviation values ​​corresponding to multiple research projects to be classified. This method can deeply explore the user's past research trajectory, accurately anchor the user's research inclinations, effectively reduce the deviation between the classification results and the user's interests, and greatly improve the personalization and fit of the classification results, making the classification more in line with the fixed preferences formed by the user's long-term research accumulation.

[0074] This implementation method integrates real-time domain dynamics and user preferences, ensuring that classification results keep pace with cutting-edge trends in the research field while preventing classifications from straying from the user's focus. This effectively improves the timeliness and domain relevance of classification results, helping users to promptly identify high-value research projects within the field. Furthermore, targeted optimization based on users' immediate research goals accurately matches different types of immediate research needs, making the classification results more aligned with users' current research tasks. This significantly enhances the practicality of the classification results, helping users quickly locate research projects that meet their current task requirements.

[0075] Figure 3 A flowchart illustrating the second research project classification method based on personalized machine learning provided in this application embodiment is shown below. Figure 3 As shown, in some implementations, in the above-mentioned S110, feature extraction is performed on the scientific research history data and scientific research stage tags to determine the user's scientific research interest profile and preference weight, including S111 to S112. S111 to S112 will be explained in detail below.

[0076] S111. Through the platform's user research database, based on the user's unique user identifier, obtain the user's research field, historical classification records, and published paper topic data.

[0077] Figure 4 A schematic diagram of the workflow of the second research project classification method based on personalized machine learning provided in the embodiments of this application is shown below. Figure 4 As shown, in this implementation, the user research database built on the platform can be used as the retrieval basis to retrieve three types of core research data associated with the user, namely the user's research field, historical classification records, and published paper topic data.

[0078] It should be noted that the user's research field can be obtained by integrating the research direction filled in by the user during registration and the field tags associated with subsequent research activities on the platform; historical classification records can be extracted to obtain all trace data of the user's past classification operations on research projects; published paper topic data can be extracted from the titles, keywords and abstracts of the published papers uploaded by the user to the platform.

[0079] For example, if a user's unique identifier is a specific code, the platform can use that code to match the computer vision field selected by the user during registration, retrieve the user's classification operation records for multiple scientific research projects within the past three years, as well as the set of thematic keywords of multiple papers published by the user.

[0080] S112. Perform One-Hot encoding and TF-IDF weight statistics on user research fields, historical classification records, and published paper topic data to determine personalized feature vectors and domain preference weight tables. Use the personalized feature vectors as research interest profiles and the domain preference weight tables as preference weights.

[0081] In this implementation, the retrieved user research field, historical classification records, and published paper topic data can be standardized. First, the discrete categorical data is converted into a computable numerical vector through One-Hot encoding. Then, the importance of different research topics in the user's research trajectory is quantified through the TF-IDF weighting method, thereby forming a personalized feature vector and a field preference weight table. The former serves as the user's research interest profile, and the latter serves as the user's preference weight.

[0082] It should be noted that One-Hot encoding can convert multiple research fields involved by a user into vectors containing only 0s and 1s, with each dimension representing a specific field. A dimension value of 1 indicates that the user is involved in that field. TF-IDF weight statistics can calculate the weight value of a topic based on the frequency of a topic appearing in the user's papers and classification records, as well as the frequency of the topic appearing in the research data of the entire platform.

[0083] For example, if a user is involved in both computer vision and natural language processing, One-Hot encoding will generate a two-dimensional vector with a corresponding dimension of 1; TF-IDF weight statistics can calculate that the user's weight for one domain topic is higher than that for the other, and form a domain preference weight table containing the weights of each topic.

[0084] It should be noted that the personalized feature vector is a digital abstraction of a user's research interests, covering feature information of all research fields and topics involved in the user; while the domain preference weight table is a quantitative presentation of the degree of attention a user pays to different research fields and topics.

[0085] For example, the personalized feature vector can contain multiple dimensions, each corresponding to a sub-topic of scientific research, and the dimension value represents the user's involvement in that topic; the domain preference weight table can list multiple sub-topics of scientific research and their corresponding weight values, clearly presenting the user's core preference domain.

[0086] This implementation method obtains users' research fields, historical classification records, and published paper topics from the platform's user research database. It then performs One-Hot encoding and TF-IDF weight statistics on this data to determine personalized feature vectors as research interest profiles and domain preference weight tables as preference weights. This allows for a more accurate depiction of users' research interest tendencies and domain preference levels, providing a more user-centric foundation for subsequent research project classification.

[0087] Figure 5 A flowchart illustrating the third method for classifying research projects based on personalized machine learning provided in this application is shown below. Figure 5 As shown, in some implementations, in S130 above, the user's research interest profile and the matching classification tags corresponding to multiple research projects to be classified are matched and filtered in multiple dimensions to obtain the matching deviation values ​​corresponding to multiple research projects to be classified, including S131 to S132. S131 to S132 will be explained in detail below.

[0088] S131. Embed the personalized feature vector and domain preference weight table into the unified feature space. Within the unified feature space, determine the domain preference weight vector corresponding to the matching classification label of the research project to be classified.

[0089] Figure 6 A schematic diagram illustrating the workflow of the third research project classification method based on personalized machine learning provided in this application embodiment is shown below. Figure 6 As shown, in this implementation, a feature space embedding algorithm can be used to map the personalized feature vector representing the user's scientific research interest profile and the domain preference weight table representing preference weights to the same standardized unified feature space, ensuring that the two are within a dimensional system in which similarity can be directly calculated. Then, in this unified feature space, the domain preference weight vector corresponding to the matching classification label of each scientific research project to be classified is located.

[0090] It should be noted that the domain preference weight vector is a digital representation of the matching classification labels of the research projects to be classified. Its dimension is consistent with the dimension of the unified feature space, and the values ​​in the vector represent the weight distribution of the classification label on each research topic.

[0091] For example, if the unified feature space has a multi-dimensional structure, and the matching classification label of a certain scientific research project to be classified is the application of deep learning in medical imaging, the corresponding domain preference weight vector will show a high weight value in the corresponding dimensions such as deep learning and medical imaging, while other dimensions will have a low value or 0.

[0092] S132. Within a unified feature space, determine the distance between the personalized feature vector and the domain preference weight vector, and the sum of the distances between the vector corresponding to the domain preference weight table and the domain preference weight vector, as the matching deviation value for multiple research projects to be classified.

[0093] In this implementation, the distance between two sets of vectors can be calculated separately in a unified feature space. The first set is the distance between the user's personalized feature vector and the domain preference weight vector of the research project to be classified. The second set is the distance between the vector corresponding to the domain preference weight table and the domain preference weight vector of the research project to be classified. The values ​​of these two distances are added together, and the result is the matching deviation value corresponding to the research project to be classified.

[0094] It should be noted that the vector distance calculation method used here can be cosine distance or Euclidean distance. The smaller the distance value, the higher the fit between the two sets of vectors; the smaller the matching deviation value, the higher the fit between the research project to be classified and the user's research interests and preferences.

[0095] For example, the distance between the personalized feature vector and the domain preference weight vector of a research project to be classified is calculated as a specific value, and the distance between the vector corresponding to the domain preference weight table and the domain preference weight vector is calculated as another specific value. The result of adding the two values ​​is the matching deviation value of the project, which can be directly used in the subsequent requirement verification and optimization process.

[0096] This implementation embeds personalized feature vectors and domain preference weight tables into a unified feature space. Within this space, the domain preference weight vectors corresponding to the matching classification labels of the research projects to be classified are determined. The distance between the personalized feature vector and the domain preference weight vector, and the sum of the distances between the vectors corresponding to the domain preference weight tables and the domain preference weight vectors are calculated as the matching deviation value. This achieves accurate quantitative matching between user interests and projects to be classified. The obtained matching deviation value can objectively reflect the degree of fit between the two, effectively improving the accuracy of subsequent classification.

[0097] This implementation method provides a precise quantitative reference for verifying and optimizing the user's current research needs and the projects to be classified, based on accurately generated research interest profiles, preference weights, and matching deviation values ​​calculated within a unified feature space. This makes the final set of demand-adaptive classifications more closely match the user's current research needs, significantly improving the adaptability and practical application value of research project classifications.

[0098] Figure 7 A flowchart illustrating the fourth research project classification method based on personalized machine learning provided in this application is shown below. Figure 7 As shown, in some implementations, the method further includes S210 to S230, which are described in detail below.

[0099] S210. Acquire users' real-time research interaction data. Using a rule-based neighbor routing unit, determine multiple peer collaboration characteristics based on personalized feature vectors, domain preference weight tables, and real-time research interaction data. This real-time research interaction data includes browsing projects, saving documents, and peer collaboration records.

[0100] Figure 8 A schematic diagram illustrating the workflow of the fourth research project classification method based on personalized machine learning provided in this application embodiment is shown below. Figure 8 As shown, this implementation method can continuously collect various real-time scientific research-related operation data generated by users within the scientific research platform, which serves as real-time scientific research interaction data and provides a data foundation for subsequent analysis of users' collaborative tendencies.

[0101] It should be noted that real-time scientific research interaction data covers dynamic scientific research operation data generated by users on the scientific research platform, which can reflect users' recent scientific research focus and collaborative behavior.

[0102] For example, browsing projects refers to the special research projects in the field of new energy materials that users view on the platform; collected literature refers to the core journal articles on perovskite battery preparation processes that users save; and peer collaboration records refer to the collaborative process of photovoltaic device R&D projects in which users and other researchers in the field jointly participate.

[0103] In this implementation, the rule-based neighbor routing unit can be invoked to integrate and analyze the user's existing personalized feature vectors, domain preference weight tables, and real-time scientific research interaction data to uncover relevant features that reflect the user's scientific research collaboration tendencies. This can uncover the user's potential tendencies in scientific research social and collaborative aspects, make up for the lack of dimensions in feature construction based solely on personal historical data, and enrich the information coverage of the feature vectors.

[0104] S220. Concatenate multiple peer collaboration features and personalized feature vectors to obtain a fused behavioral feature vector. Obtain incremental data on user research interactions. Iteratively update the fused behavioral feature vector and incremental research interaction data using sliding window statistics to obtain an updated fused behavioral feature vector. Specifically, the peer collaboration features and personalized feature vectors are concatenated with a 7:3 weight.

[0105] In this implementation, multiple extracted peer collaboration features and the user's existing personalized feature vectors can be combined and spliced ​​to generate a fused feature vector that includes the user's personal research inclinations and collaboration inclinations.

[0106] It should be noted that the feature splicing process can allocate the proportion of peer collaboration features and personalized feature vectors according to a preset weight ratio to ensure that the fused vector can reasonably reflect the influence of the two types of features.

[0107] For example, peer collaboration feature data and personalized feature vector data can be combined by weighting the corresponding dimensions according to a 7:3 ratio to generate a complete fused behavioral feature vector.

[0108] In this implementation, newly generated scientific research interaction data from users within a set time period can be collected to provide the latest data support for the iterative update of feature vectors.

[0109] It should be noted that incremental research interaction data refers to the research operation data added by users within the feature vector update cycle, which can reflect the latest changes in users' research behavior.

[0110] For example, incremental data on scientific research interactions may include energy storage technology projects that users have recently viewed, newly collected solid electrolyte research literature, and newly established collaboration records with researchers in the field of energy storage.

[0111] In this implementation, a sliding window statistical method can be used to dynamically integrate the fused behavioral feature vector and incremental scientific research interaction data, thereby enabling iterative updates of the feature vector to better match the user's current scientific research status. The dynamic update mechanism can capture the latest changes in the user's scientific research behavior in real time, avoid feature solidification, and improve the fit of the features to the user's current scientific research status.

[0112] It should be noted that the sliding window statistics can be set to a fixed time window range, retaining only the latest scientific research interaction data within that range for feature updates, ensuring that the feature vector can reflect the user's recent scientific research behavior in real time.

[0113] For example, a 7-day time range can be set for the sliding window. Each time an update is performed, old interaction data from 7 days ago is removed, and incremental scientific research interaction data from the past 7 days is added. The values ​​of each dimension of the fusion behavior feature vector are adjusted to obtain the updated fusion behavior feature vector.

[0114] S230. Map the updated fusion behavior feature vector and the research project attribute data of the research projects to be classified to the same feature space and standardize them. Through cosine similarity matching, determine the user-project matching values ​​corresponding to multiple updated fusion behavior feature vectors and the research projects to be classified. Determine multiple target user-project matching values ​​that are greater than or equal to the preset user-project matching values, and use the target updated fusion behavior feature vectors corresponding to the multiple target user-project matching values ​​as research interest profiles.

[0115] In this implementation, the updated fusion behavior feature vector and the attribute data of the scientific research project to be classified can be converted into a unified feature space, and the two types of data can be standardized to eliminate the impact of data dimensional differences on the matching results.

[0116] In this implementation, the similarity between the updated fused behavioral feature vector and the attribute data vector of the scientific research project to be classified within a unified feature space can be calculated using the cosine similarity algorithm, thus obtaining a matching value that reflects the degree of fit between the two.

[0117] It should be noted that cosine similarity matching calculates the cosine of the angle between two vectors to quantitatively assess the similarity between user characteristics and research project attributes; the higher the value, the higher the degree of fit.

[0118] For example, the updated fusion behavior feature vector and the hydrogen energy utilization scientific research project attribute data vector to be classified can be used to calculate cosine similarity to obtain a specific user-project matching value, which reflects the degree of fit between the user and the hydrogen energy project.

[0119] In this implementation, a matching value threshold can be set to filter out target matching values ​​that meet the threshold requirements. The corresponding updated fusion behavior feature vector is used as the user's research interest profile, providing accurate feature support for subsequent research project classification. The research interest profile can be optimized based on real-time matching results, making it more accurately suited to the research projects to be classified and improving the accuracy of subsequent classification matching.

[0120] It should be noted that the preset user-project matching value is a pre-set fit threshold used to filter out feature vectors that can effectively represent the user's current research interests.

[0121] For example, a specific preset user-project matching value can be set, and target feature vectors that match the value can be filtered out and updated as the user's research interest profile.

[0122] This implementation method acquires users' real-time scientific research interaction data, which includes browsing projects, collecting literature, and peer collaboration records. Through a rule-based neighbor routing unit, combined with personalized feature vectors, domain preference weight tables, and real-time scientific research interaction data, multiple peer collaboration features are determined. By mining users' potential tendencies in scientific research social interaction and collaboration, the method compensates for the lack of dimensions in feature construction based solely on personal historical data and enriches the information coverage of feature vectors.

[0123] This implementation method concatenates multiple peer collaboration features and personalized feature vector weights to obtain a fusion behavior feature vector, acquires incremental data of user research interactions, and iteratively updates the fusion behavior feature vector and incremental data of research interactions through sliding window statistics to obtain an updated fusion behavior feature vector. The dynamic update mechanism can capture the latest changes in user research behavior in real time, avoid feature solidification, and improve the fit of features to the user's current research status.

[0124] This implementation maps the updated fusion behavior feature vector and the research project attribute data of the research projects to be classified to the same feature space and standardizes them. Cosine similarity matching is used to determine the user-project matching values ​​corresponding to multiple updated fusion behavior feature vectors and the research projects to be classified. Target user-project matching values ​​that are greater than or equal to the preset user-project matching values ​​are selected, and their corresponding target updated fusion behavior feature vectors are used as research interest profiles. The research interest profiles are optimized based on real-time matching results to accurately match the research projects to be classified and improve the accuracy of subsequent classification matching.

[0125] Figure 9 A flowchart illustrating the fifth method for classifying research projects based on personalized machine learning, as provided in the embodiments of this application, is shown below. Figure 9 As shown, in some implementations, in the above-mentioned S210, the rule-based neighbor routing unit determines multiple peer collaboration features based on personalized feature vectors, domain preference weight tables and real-time scientific research interaction data, including S211 to S212. S211 to S212 will be explained in detail below.

[0126] S211. Determine the user's core research field through a domain preference weight table. Using preset rules, identify multiple target real-time research interaction data points from real-time research interaction data that match the core research field, meet collaboration frequency standards, and have valid collaboration types.

[0127] Figure 10 A schematic diagram illustrating the workflow of the fifth research project classification method based on personalized machine learning provided in this application embodiment is shown below. Figure 10As shown, in this implementation, the user's core research field can be identified by using a domain preference weight table. Specifically, multiple target real-time scientific research interaction data that match the core research field, meet the collaboration frequency standard, and have effective collaboration type can be selected from the real-time scientific research interaction data based on preset rules.

[0128] It should be noted that when determining a user's core research area through the domain preference weight table, the top few research areas with the highest weight in the table can be extracted and designated as the user's core research area. These areas are the directions in which the user has devoted the most effort and attention in past scientific research activities.

[0129] For example, if the combined weight of "Artificial Intelligence", "Machine Learning" and "Data Mining" in the domain preference weight table exceeds 70% of the overall weight, and the weight of other domains is relatively low, then these three domains can be identified as the user's core research domains.

[0130] It should be noted that the preset rules can cover three screening conditions. The first is that the research field corresponding to the interactive data must fall within the scope of the core research field; the second is that the frequency of collaboration with the object reaches the set qualification standard; and the third is that the type of collaboration belongs to the type with substantial scientific research value.

[0131] For example, collaboration records with peers in the core research field are filtered out from the user's real-time scientific research interaction data. The collaboration frequency reaches the set qualified value, and the collaboration type is joint application for scientific research projects or co-authored academic papers. These data are identified as target real-time scientific research interaction data.

[0132] S212. Determine the collaboration weight based on the collaboration frequency corresponding to real-time scientific research interaction data of multiple targets. Multiply the domain preference weight in the domain preference weight table by the collaboration weight to obtain the modified domain preference weight table, and use the modified domain preference weight table as the collaboration feature of multiple peers.

[0133] In this implementation, the collaboration weight can be determined based on the collaboration frequency corresponding to the real-time scientific research interaction data of multiple targets. Furthermore, the domain preference weight in the domain preference weight table can be multiplied by the collaboration weight to obtain a modified domain preference weight table, which can then be used as a feature of collaboration among multiple peers.

[0134] It should be noted that the collaboration weight can be determined based on the gradient division of collaboration frequency to set corresponding weight values. The higher the collaboration frequency, the greater the corresponding collaboration weight, thereby quantifying the impact of collaboration behavior on users' research interests.

[0135] For example, corresponding weights are set for different collaboration frequency ranges. The higher the frequency of the range, the greater the corresponding weight value. If the collaboration frequency of a certain target's real-time scientific research interaction data is in the high-frequency range, then a corresponding high collaboration weight is assigned.

[0136] It should be noted that each domain preference weight in the domain preference weight table is multiplied by its corresponding collaboration weight to obtain the corrected domain preference weight. These corrected weights are then integrated to form the corrected domain preference weight table, which reflects the changes in users' research domain preferences caused by peer collaboration.

[0137] For example, the weight of a core research area in the original domain preference weight table is a fixed value, and the corresponding collaboration weight is a set value. The two are multiplied to obtain the corrected weight of the area. After processing all areas in turn, they are integrated to form a corrected domain preference weight table, which is used as the peer collaboration feature.

[0138] This implementation method determines the user's core research field through a domain preference weight table. Based on preset rules, it filters out multiple target real-time scientific research interaction data from real-time scientific research interaction data that match the core research field, meet the collaboration frequency standard, and have effective collaboration types. This ensures that the selected interaction data is closely related to the user's core research direction and has effective collaboration attributes, effectively filtering interaction data that is irrelevant to the core research field or whose collaboration quality is substandard. This provides accurate and high-quality data support for the subsequent generation of peer collaboration features, and strengthens the accurate mapping of peer collaboration features to the user's scientific research interests.

[0139] This implementation determines collaboration weights based on the collaboration frequency corresponding to real-time scientific research interaction data from multiple targets. The domain preference weights in the domain preference weight table are multiplied by the collaboration weights to obtain a modified domain preference weight table. This modified domain preference weight table is then used as a feature of multiple peer collaborations. By quantifying the impact of collaboration on users' research interests through collaboration frequency, the domain preference weights are dynamically modified. This makes the peer collaboration features more closely match the actual collaboration intensity of users, more accurately reflects changes in users' research interests caused by collaboration, and optimizes the accuracy of subsequent updates to the research interest profile.

[0140] This approach first anchors the user's core research area to filter effective interaction data, then assigns collaboration weights based on collaboration frequency, and generates peer collaboration features by adjusting the domain preference weight table through weight adjustment. This integrates the dual influence of the user's core research preferences and effective collaborative behavior, avoiding interference from invalid or weakly correlated collaboration data on the accuracy of peer collaboration features. This ensures that peer collaboration features not only align with the user's core research direction but also reflect the impact of collaborative behavior on research interests, providing a more accurate feature basis for the subsequent generation of fusion behavior feature vectors and improving the reliability of user-project matching values.

[0141] Figure 11A flowchart illustrating the sixth method for classifying research projects based on personalized machine learning, as provided in the embodiments of this application, is shown below. Figure 11 As shown, in some implementations, in the above-mentioned S230, multiple target user-item matching values ​​that are greater than or equal to the preset user-item matching value are determined, including S231 to S232. S231 to S232 will be explained in detail below.

[0142] S231. Through the platform's user research database, obtain the user's historical classification feedback data based on the user's unique identifier. Determine the project classification screening threshold based on the historical classification feedback data.

[0143] Figure 12 A schematic diagram illustrating the workflow of the sixth research project classification method based on personalized machine learning provided in this application embodiment is shown below. Figure 12 As shown, in this implementation, the platform's user research database can be used to retrieve the user's historical categorized feedback data based on the user's unique identifier.

[0144] It should be noted that the historical classification feedback data includes users' past approval and rejection tags for the system's recommended scientific research project classifications, as well as records of manual adjustments to the classification results.

[0145] For example, a user may have previously marked most of the projects in a specific research direction category pushed by the system as approved, and manually adjusted some projects in another direction category. These data are all historical category feedback data.

[0146] In this implementation, the obtained historical classification feedback data can be used to determine the project classification screening threshold.

[0147] It should be noted that historical classification feedback data can be mapped to corresponding project classification filtering thresholds through a preset experience value table. The experience value table contains threshold ranges corresponding to different feedback ratios.

[0148] For example, when the percentage of users who approve of a certain category of results falls within a specific range, the corresponding item category screening threshold can be obtained by matching the empirical value table.

[0149] S232. Determine the research stage adjustment coefficient based on the research stage label. Determine the product of the project classification screening threshold and the research stage adjustment coefficient as the stage project classification screening threshold. Use the stage project classification screening threshold as the preset user-project matching value, and determine multiple target user-project matching values ​​that are greater than or equal to the preset user-project matching value.

[0150] In this implementation, the corresponding research stage adjustment coefficient can be determined based on the user's research stage tag.

[0151] It should be noted that the coefficients can be adjusted by matching the corresponding research stage labels using a preset empirical value table. The empirical value table covers the coefficient ranges corresponding to different research stages.

[0152] For example, users in the core stage of project application are assigned a research stage adjustment coefficient that matches the needs of that stage; users in the in-depth literature review stage are assigned a different coefficient value.

[0153] In this implementation, the product of the project classification screening threshold and the scientific research stage adjustment coefficient can be calculated, and the result of this product can be used as the stage project classification screening threshold.

[0154] It should be noted that by multiplying the basic threshold determined based on historical feedback with the adjustment coefficient adapted to the current research stage, the resulting stage project classification screening threshold combines both user historical preferences and current stage characteristics.

[0155] For example, the determined project classification screening threshold and the scientific research stage adjustment coefficient are multiplied together, and the result is the stage project classification screening threshold.

[0156] In this implementation, the stage project classification filtering threshold can be set to a preset user-project matching value, and multiple target user-project matching values ​​greater than or equal to this value can be filtered out.

[0157] In this implementation, the matching results can be filtered by dynamically adjusted preset values, so that the filtered target user-project matching values ​​are more in line with the user's historical classification habits and current research needs, and reduce matching items that do not meet the user's expectations.

[0158] This implementation method utilizes the platform's user research database to obtain users' historical classification feedback data based on their unique user identifiers. This data is used to determine the project classification screening threshold. Then, a research stage adjustment coefficient is determined by combining the research stage tag. The product of the two is used as the stage project classification screening threshold. Multiple target user-project matching values ​​greater than or equal to this value are selected based on this standard. This makes the preset screening value more in line with users' past classification preferences, effectively improving the personalization of the matching results and reducing matching items that do not conform to users' habits.

[0159] This implementation method first obtains the user's research stage tag when determining the preset user-project matching value, determines the corresponding research stage adjustment coefficient, and then multiplies it with the project classification screening threshold obtained from historical classification feedback data to obtain the stage project classification screening threshold. The target user-project matching value is then screened using this threshold, so that the screening criteria are adapted to the user's current research stage, meet the research needs of different stages, make the matching results more in line with the user's current research pace, and improve the accuracy of classification.

[0160] This implementation method eliminates the use of fixed preset user-project matching values. Instead, it combines a basic threshold determined by historical user feedback data with dynamic stage project classification filtering thresholds adjusted based on research stage tags. This allows for the selection of target user-project matching values ​​that meet the criteria. This avoids fixed thresholds and enables the filtering standards to be dynamically adjusted based on historical user feedback and the current research stage, further optimizing the classification and matching results of research projects and improving the user experience.

[0161] Figure 13 A flowchart illustrating the seventh method for classifying research projects based on personalized machine learning, as provided in the embodiments of this application, is shown below. Figure 13 As shown, in some implementations, the domain heat labeling unit is trained through S121 to S122. S121 to S122 will be explained in detail below.

[0162] S121. Map the sample research interest profile, sample preference weight, sample real-time domain hot data, research project attribute data of multiple sample research projects to be classified, sample real-time labels of sample-labeled research projects, and sample popularity values ​​to a unified vector space.

[0163] In this implementation, the sample research interest profile, sample preference weight, sample real-time domain hot data, research project attribute data of multiple sample research projects to be classified, sample real-time labels of sample-labeled research projects and sample popularity values ​​can be mapped to a unified vector space, providing dimensionally consistent basic data for subsequent model training and eliminating the feature heterogeneity of multi-source data.

[0164] For example, the sample research interest profile can be represented by a vector characterizing the interests of researchers in a certain biomedical field; the sample preference weight can be represented by the weight distribution table of the researcher's interest in the tumor immunotherapy subfield; the sample real-time field hotspot data can be represented by a set of hot research directions in the biomedical field during a certain period; the research project attribute data of multiple sample research projects to be classified can be represented by the theme, research content, etc. of the CAR-T cell therapy and antibody drug projects to be classified; the sample real-time label for the research project can be "solid tumor CAR-T therapy", and the sample popularity value can be represented by the domain attention quantification value of the corresponding research project.

[0165] S122. A domain popularity labeling unit is trained using sample research interest profiles, sample preference weights, real-time domain hotspot data, research project attribute data of multiple research projects to be classified, real-time labels of multiple sample labeled research projects, and sample popularity values. The domain popularity labeling unit includes an input layer, a feature alignment layer, a dual-branch matching layer, and a fusion calculation layer, arranged sequentially. The dual-branch matching layer includes a user interest matching branch layer and a domain hotspot association branch layer. The user interest matching branch layer determines the basic matching degree between the research project attribute vector corresponding to the research project attribute data and the research interest vector corresponding to the research interest profile, and determines the product of the basic matching degree and the preference weight vector corresponding to the preference weight as the weighted interest matching degree. The domain hotspot association branch layer determines the basic correlation degree between the research project attribute vector and the real-time domain hotspot vector corresponding to the real-time domain hotspot data, and inputs the basic correlation degree and timeliness coefficient into a time decay weighting module, which outputs a corrected hotspot correlation degree. The weighted interest matching degree, the corrected hotspot correlation degree, and the vectors corresponding to the real-time labels and sample popularity values ​​of multiple sample labeled research projects are input into the fusion calculation layer.

[0166] In this implementation, various types of sample scientific research data mapped to a unified vector space can be used to train domain popularity labeling units. These units learn the correlation between scientific research interests, domain hotspots, and labeled scientific research projects, providing reliable model support for the subsequent generation of accurate labeled scientific research projects.

[0167] In this implementation, a domain heat labeling unit can be constructed, which includes an input layer, a feature alignment layer, a dual-branch matching layer, and a fusion calculation layer. The dual-branch matching layer is configured with a user interest matching branch layer and a domain hotspot association branch layer. Through the collaborative operation of the multi-layer structure, the effective fusion and accurate calculation of multi-dimensional scientific research data can be achieved.

[0168] It should be noted that the feature alignment layer can align features from different sources, eliminate interference caused by differences in different data dimensions, and ensure that all types of data participate in subsequent matching calculations under the same dimension.

[0169] In this implementation, the user interest matching branch layer can first determine the basic matching degree between the scientific research project attribute vector corresponding to the scientific research project attribute data and the scientific research interest vector corresponding to the scientific research interest profile. Then, the basic matching degree and the preference weight vector corresponding to the preference weight are multiplied to obtain the weighted interest matching degree, thereby reflecting the degree of influence of the user's different scientific research field preferences on the matching results.

[0170] In this implementation, the basic correlation between the research project attribute vector and the real-time domain hotspot vector corresponding to the real-time domain hotspot data can be determined through the domain hotspot association branch layer. Then, the basic correlation and timeliness coefficient are input into the time decay weighting module, which outputs the corrected hotspot correlation, thereby dynamically considering the impact of the timeliness changes of domain hotspots on the correlation.

[0171] It should be noted that the time decay weighting module can adjust the weights according to the generation time of hot data. The longer the hot data is generated, the lower its corresponding correction correlation, thereby avoiding the interference of outdated hot data on the annotation results.

[0172] In this implementation, the weighted interest matching degree, the corrected hotspot correlation degree, and the vectors corresponding to the real-time labels and sample popularity values ​​of multiple sample-annotated scientific research projects can be input into the fusion calculation layer. The training of the domain popularity annotation unit is completed through the fusion operation within the layer, and the multi-dimensional data is integrated to output accurate annotation results.

[0173] This implementation maps sample research interest profiles, sample preference weights, real-time domain hotspot data, research project attribute data of multiple research projects to be classified, real-time labels of sample-labeled research projects, and sample popularity values ​​to a unified vector space. It then trains a domain popularity labeling unit. The dual-branch matching layer includes a user interest matching branch and a domain hotspot association branch. The former calculates the basic matching degree between the research project attribute vector and the research interest vector, multiplying it by the preference weight vector to obtain a weighted interest matching degree. The latter calculates the basic association degree and obtains a corrected hotspot association degree through a time decay weighting module. Finally, a fusion calculation layer integrates the data to complete the training, taking into account both user research interest weights and the timeliness of domain hotspots. This makes the labeled research projects more aligned with needs and real-time trends, improving classification accuracy.

[0174] In this implementation, the domain hotspot association branch layer of the domain heat labeling unit calculates the basic correlation between the research project attribute vector and the real-time domain hotspot vector. This correlation, along with the timeliness coefficient, is then input into the time decay weighting module to obtain the corrected hotspot correlation. Finally, this correlation is input into the fusion calculation layer along with the weighted interest matching degree, the real-time sample label, and the sample heat value vector to complete the training. This approach can dynamically consider the timeliness changes of domain hotspots, avoiding interference from outdated hotspots on the labeling results. At the same time, it integrates the dual dimensions of user interest and hotspots, making the real-time labels and heat values ​​of labeled research projects more accurate and improving the fit between subsequent projects to be classified and user needs.

[0175] Figure 14 A flowchart illustrating the eighth method for classifying research projects based on personalized machine learning, as provided in the embodiments of this application, is shown below. Figure 14As shown, in some implementations, in S130 above, the user's current research needs and the matching deviation values ​​corresponding to multiple research projects to be classified are optimized by demand verification to obtain a demand adaptation classification set of multiple research projects to be classified that are adapted to the current research needs. The demand adaptation classification set includes multiple demand adaptation classifications that are adapted to the current research needs, including S310 to S320. S310 to S320 will be explained in detail below.

[0176] S310. Obtain the feasibility weight, the popularity weight of the first field, and the first matching deviation weight. When the current research need is a project application, obtain the feasibility indicators and the popularity of the first field corresponding to the current research need.

[0177] In this implementation, we can obtain the feasibility weight, the popularity weight of the first field, and the first matching deviation weight. In the scenario where the current research need is a project application, we can obtain the feasibility indicators and the popularity of the first field corresponding to the need, providing multi-dimensional basic parameters for subsequent comprehensive evaluation.

[0178] It should be noted that the weights for feasibility, popularity in the first field, and first matching deviation are weight parameters set according to the core needs of the project application scenario, and are used to balance the influence of each evaluation dimension.

[0179] For example, the weights for feasibility of implementation, popularity of the first field, and first matching deviation can be set according to the actual needs of the project application. For project application scenarios that focus on the transformation of results, a higher weight for feasibility of implementation can be set; for scenarios that focus on cutting-edge exploration, a higher weight for popularity of the first field can be set.

[0180] It should be noted that the feasibility indicators and popularity of the primary research field corresponding to current research needs can be obtained through an experience value table, which is based on historical data from a large number of research project applications.

[0181] For example, by using the experience value table, when the current research need is to apply for a research project, the feasibility indicator can be obtained. This indicator reflects the potential for the transformation of research results and resource suitability of the research project to be classified. At the same time, the popularity of the first field can be obtained. This indicator reflects the current research popularity and attention received in the field to which the research project to be classified belongs.

[0182] S320. Determine the sum of the product of the feasibility indicator and the feasibility weight, and the product of the first domain popularity and the first domain popularity weight, as the feasibility evaluation indicator. Determine the product of the matching deviation value and the first matching deviation weight, as the matching deviation evaluation indicator. Determine the difference between the feasibility evaluation indicator and the matching deviation evaluation indicator, as the comprehensive evaluation indicator for multiple research projects to be classified. Determine multiple research projects to be classified whose comprehensive evaluation indicator is greater than or equal to the preset comprehensive evaluation indicator, thus obtaining a demand adaptation classification set for multiple research projects to be classified to meet the current research needs.

[0183] In this implementation method, the product of the feasibility indicator and the feasibility weight can be calculated, and then the product of the popularity of the first field and the popularity weight of the first field can be calculated. The results of these two products are added together to obtain the feasibility evaluation index. This index comprehensively reflects the overall performance of the scientific research project to be classified in terms of feasibility and popularity of the field.

[0184] In this implementation, the product of the matching deviation value and the first matching deviation weight can be calculated to obtain the matching deviation evaluation index. This index quantifies the weight ratio of the degree of deviation between the scientific research project to be classified and the user's scientific research interests.

[0185] In this implementation, the difference between the feasibility assessment index and the matching deviation assessment index can be calculated, and this difference can be used as a comprehensive assessment index for multiple research projects to be classified. This index comprehensively considers the balance between the feasibility of the research projects to be classified, the popularity of the field, and the matching deviation of user interests.

[0186] It should be noted that the comprehensive evaluation index is obtained by subtracting the matching deviation evaluation index from the feasibility evaluation index. This calculation method can highlight the positive contribution of feasibility and popularity in the field, while weakening the negative impact of matching deviation, which is in line with the core needs of project application.

[0187] For example, if the feasibility assessment indicator for a certain research project to be classified is a quantitative value of the corresponding dimension, and the matching deviation assessment indicator is a quantitative value of the corresponding dimension, the difference between the two is used as the comprehensive assessment indicator of the project. The larger the difference, the more the advantages of the project in terms of implementation and popularity can offset the deviation from user interests.

[0188] In this implementation method, multiple research projects to be classified can be identified whose comprehensive evaluation indicators are greater than or equal to the preset comprehensive evaluation indicators. These projects are then integrated to obtain a set of demand-adaptive classifications that meet the current project application requirements. This set contains multiple demand-adaptive classifications that meet the requirements.

[0189] It should be noted that the preset comprehensive evaluation indicators are thresholds set according to the overall requirements of the project application, and are used to screen research projects that meet the basic evaluation criteria for classification.

[0190] For example, if the preset comprehensive evaluation index is the quantitative threshold of the corresponding dimension, when the comprehensive evaluation index of the research project to be classified is greater than or equal to the threshold, it will be included in the demand adaptation classification set. The projects in this set not only have good feasibility and popularity in the field, but can also match the user's research interests to a certain extent and adapt to the needs of project application.

[0191] This implementation method, when the current research need is project application, obtains the feasibility weight, the first field popularity weight, and the first matching deviation weight. It then obtains the corresponding feasibility index and the first field popularity. The product of the feasibility index and the feasibility weight, plus the product of the first field popularity and its weight, yields the feasibility evaluation index. The product of the matching deviation value and the first matching deviation weight is then calculated to obtain the matching deviation evaluation index. Subtracting the matching deviation evaluation index from the feasibility evaluation index yields the comprehensive evaluation index. Research projects meeting the comprehensive evaluation index are then selected to form a demand-fitting classification set. This more accurately selects suitable research projects based on the core needs of project applications for feasibility and field popularity, improving the practicality of the classification results in project application scenarios.

[0192] This approach achieves a reasonable balance among the various evaluation dimensions through weight allocation, taking into account both the user's own research interests and the requirements of project application for practical value and field popularity. It avoids the one-sidedness of single-dimensional evaluation and improves the comprehensiveness of the classification results. Compared with the original method that only relies on matching deviation value and demand verification, it can more accurately capture the core needs of project application, filter out projects with small matching deviation but insufficient practicality or popularity, and improve the accuracy of demand adaptation.

[0193] Figure 15 A flowchart illustrating the ninth method for classifying research projects based on personalized machine learning, as provided in this application embodiment, is shown below. Figure 15 As shown, in some implementations, in S130 above, the user's current research needs and the matching deviation values ​​corresponding to multiple research projects to be classified are optimized by demand verification to obtain a demand adaptation classification set of multiple research projects to be classified that are adapted to the current research needs. The demand adaptation classification set includes multiple demand adaptation classifications that are adapted to the current research needs, and also includes S330 to S340. S330 to S340 will be explained in detail below.

[0194] S330: Obtain academic relevance weights, second-domain popularity weights, and second-matching deviation weights. When the current research need is literature review, obtain the academic relevance indicators and second-domain popularity corresponding to the current research need.

[0195] In this implementation, academic relevance weights, second field popularity weights, and second matching deviation weights can be pre-set and obtained for literature research needs assessment. At the same time, when the current research need is identified as literature research, the core assessment indicator data for the corresponding scenario can be obtained.

[0196] It should be noted that the academic relevance weight, the second field popularity weight, and the second matching deviation weight are used to assign different influence percentages to the three evaluation dimensions of academic relevance, field popularity, and user interest matching deviation in the literature research scenario.

[0197] For example, an empirical value table can be used to set academic relevance weight, second field popularity weight, and second matching deviation weight, and the numerical proportion of each weight can be adjusted according to the needs of the literature research scenario.

[0198] It should be noted that the academic-related indicators are used to measure the depth of academic content and citation relevance of the research projects to be classified, while the second-domain popularity is used to measure the current level of attention in the field to which the project belongs.

[0199] For example, the academic-related indicators and the popularity of the second field corresponding to the current research needs can be obtained through an empirical value table. The academic-related indicators are determined based on the number of literature citations and core journal publications of the research projects to be classified, and the popularity of the second field is determined based on the number of recent project releases and the frequency of academic discussions within the field.

[0200] S340. Determine the sum of the product of academic relevance indicators and academic relevance weights, and the product of the second-domain popularity and the second-domain popularity weights, as the academic relevance evaluation indicator. Determine the product of the matching deviation value and the second matching deviation weights, as the matching deviation evaluation indicator. Determine the difference between the academic relevance evaluation indicator and the matching deviation evaluation indicator, as the comprehensive evaluation indicator for multiple research projects to be classified. Identify multiple research projects to be classified whose comprehensive evaluation indicator is greater than or equal to the preset comprehensive evaluation indicator, thus obtaining a set of demand-fit classifications for multiple research projects to be classified to meet current research needs.

[0201] In this implementation, the academic-related indicators and academic-related weights can be multiplied together, and the popularity of the second field and the popularity weights of the second field can be multiplied together. The two product results are then added together to obtain the academic-related evaluation index, which comprehensively reflects the academic value and field popularity value of the research project to be classified.

[0202] For example, the matching deviation value corresponding to the research project to be classified and the weight of the second matching deviation can be multiplied to obtain the matching deviation evaluation index, which reflects the degree of deviation between the research project to be classified and the user's research interests.

[0203] In this implementation, the calculated academic-related evaluation index can be subtracted from the matching deviation evaluation index, and the difference obtained is the comprehensive evaluation index of the research project to be classified. The higher the value of this index, the more the project is suitable for the user's literature research needs.

[0204] In this implementation method, a unified comprehensive evaluation standard can be set to screen out research projects that meet the standard and form a classification set that adapts to the current literature survey needs.

[0205] For example, a qualified threshold for the comprehensive evaluation index can be preset, and research projects to be classified that have a comprehensive evaluation index greater than or equal to the threshold can be screened out. These projects can then be sorted according to their respective categories to obtain a set of demand-fitting categories. All categories in the set are categories that are suitable for users' literature research needs.

[0206] This implementation method, targeting current research needs in literature surveys, obtains academic relevance weights, secondary field popularity weights, and secondary matching deviation weights. It then obtains corresponding academic relevance indicators and secondary field popularity. The sum of the products of the academic relevance indicators and academic relevance weights, and the products of the secondary field popularity and secondary field popularity weights, yields the academic relevance evaluation index. Simultaneously, the matching deviation value is calculated, and its product with the secondary matching deviation weights yields the matching deviation evaluation index. The difference between the two is used as a comprehensive evaluation index. Research projects with a comprehensive evaluation index greater than or equal to the preset comprehensive evaluation index are selected to form a demand-fitting classification set. This set specifically addresses the core needs of literature surveys, improving the fit between the research projects to be classified and the user's literature survey needs, making the selection results more consistent with the user's actual academic research requirements.

[0207] This approach constructs a multi-weighted comprehensive evaluation system for literature research scenarios, consisting of academic relevance weight, second-domain popularity weight, and second-matching deviation weight. By calculating the weighted indicators for each dimension of academic relevance, domain popularity, and matching deviation separately, and then integrating them to obtain a comprehensive evaluation indicator, this system serves as the core basis for screening research projects to be classified. This breaks through the limitations of single-dimensional evaluation, comprehensively considers multiple dimensions, and makes the screening results of research projects to be classified more scientific and comprehensive, effectively reducing matching bias.

[0208] This implementation method supplements the dedicated adaptation process for literature research needs. By setting corresponding weights and evaluation indicators, it completes the screening of research projects to be classified under this need, forming a complete adaptation system covering two current research needs: project application and literature research. This fills the gap in literature research need adaptation, achieves comprehensive coverage of the core needs of users at different research stages, and greatly improves the completeness and practicality of the overall research project classification service.

[0209] In some implementation methods, the weight for feasibility is 0.3, the weight for popularity in the first field is 0.2, and the weight for the first matching deviation is 0.5. The weight for academic relevance is 0.4, the weight for popularity in the second field is 0.1, and the weight for the second matching deviation is 0.5.

[0210] In this implementation, a fixed weighting scheme can be adopted for the current scientific research needs of project application. The weighting of feasibility is set to 0.3, the weighting of the first field popularity is set to 0.2, and the weighting of the first matching deviation is set to 0.5. This weighting allocation can prioritize the consideration of the matching degree of users' scientific research interests, while taking into account the evaluation dimensions of project feasibility and field popularity.

[0211] In this implementation method, based on the above weight configuration, the feasibility assessment index and the matching deviation assessment index can be calculated respectively. Then, the comprehensive assessment index can be obtained by the difference between the two. Using this as the standard, scientific research projects that meet the needs of project application can be screened, which can make the classification results more in line with the dual needs of project application for feasibility and interest alignment.

[0212] In this implementation, a differentiated fixed weight configuration scheme can be adopted according to the current scientific research needs of literature research. The academic relevance weight is set to 0.4, the second field popularity weight is set to 0.1, and the second matching deviation weight is set to 0.5. This weight allocation can highlight the core position of academic relevance in literature research, while balancing the influence of field popularity and user interest matching.

[0213] In this implementation, academic-related evaluation indicators and matching deviation evaluation indicators can be calculated based on the weight configuration. The difference between the two can be used to obtain a comprehensive evaluation indicator, which can then be used to screen research projects that meet the needs of literature research. This allows the classification results to better align with the core academic value requirements of literature research.

[0214] This implementation method adopts differentiated weight configuration schemes for different current scientific research needs, corresponding to two scenarios: project application and literature review. By clarifying the weight allocation logic, it standardizes the calculation process of comprehensive evaluation indicators, unifies the criteria for screening the classification set that matches the needs, and unifies the classification evaluation rules under different scientific research needs through differentiated configuration. This avoids classification deviations caused by ambiguous weight settings, improves the consistency and reliability of classification adaptation under various scientific research needs, and provides users with scientific research project classification services that are more in line with their own needs.

[0215] This application also provides a scientific research project classification device based on personalized machine learning, including a unit for implementing the method described above.

[0216] Figure 16 A schematic diagram of the logical structure of a scientific research project classification method system based on personalized machine learning, as provided in the embodiments of this application, is shown below. Figure 16 As shown, the system 1 of this embodiment includes a processing unit 11, a storage unit 12, and a transceiver unit 13. The processing unit 11 is used to process data, the storage unit 12 is used to store data, and the transceiver unit 13 is used to send and receive data. The processing unit 11, the storage unit 12, and the transceiver unit 13 cooperate with each other to implement the above-described method. The beneficial effects of the embodiments of this application have been described in the above-described method and will not be repeated here.

[0217] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0218] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0219] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0220] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0221] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0222] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0223] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0224] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A scientific research project classification method based on personalized machine learning, characterized in that, The method includes: Acquire users' research history data and research stage tags; extract features from the research history data and research stage tags to determine the user's research interest profile and preference weights; among which, the research history data includes projects followed, classification habits, and research results; The system acquires real-time domain hotspot data and research project attribute data for multiple research projects to be classified; using domain popularity annotation units, based on research interest profiles, preference weights, real-time domain hotspot data, and research project attribute data for multiple research projects to be classified, it determines the annotated research projects corresponding to multiple research projects to be classified; using a preset project classification system, it classifies and evaluates the multiple annotated research projects to obtain matching classification tags for multiple research projects to be classified; among them, the annotated research projects have real-time tags and popularity values; The system performs multi-dimensional matching and filtering on user research interest profiles and matching category tags corresponding to multiple unclassified research projects to obtain matching deviation values ​​for multiple unclassified research projects; it also obtains the user's current research needs; and performs requirement verification and optimization on the user's current research needs and matching deviation values ​​corresponding to multiple unclassified research projects to obtain a set of requirement adaptation categories that adapt to the current research needs. The set of requirement adaptation categories includes multiple requirement adaptation categories that adapt to the current research needs; among which, the current research needs include project application and literature review. The matching deviation values ​​of the user's current research needs and multiple unclassified research projects are used for requirement verification and optimization. This results in a requirement adaptation category set that adapts multiple unclassified research projects to the current research needs. The requirement adaptation category set includes multiple requirement adaptation categories that adapt to the current research needs, including: Obtain the feasibility weight, the popularity weight of the first field, and the first matching deviation weight; when the current research need is a project application, obtain the feasibility indicators and the popularity of the first field corresponding to the current research need; The feasibility evaluation index is determined by multiplying the feasibility index and its weight, and by summing the product of the first field popularity and its weight. The matching deviation value and its weight are used as the matching deviation evaluation index. The difference between the feasibility evaluation index and the matching deviation evaluation index is used as the comprehensive evaluation index for multiple research projects to be classified. Multiple research projects whose comprehensive evaluation index is greater than or equal to the preset comprehensive evaluation index are identified, resulting in a set of demand-fit classifications for multiple research projects to be classified to meet current research needs.

2. The method according to claim 1, characterized in that, Feature extraction is performed on historical research data and research stage tags to determine the user's research interest profile and preference weights, including: By leveraging the platform's user research database and based on each user's unique identifier, we can obtain data on the user's research field, historical classification records, and published paper topics. One-Hot encoding and TF-IDF weight statistics are performed on user research fields, historical classification records, and published paper topic data to determine personalized feature vectors and field preference weight tables. The personalized feature vectors are used as research interest profiles, and the field preference weight tables are used as preference weights. Multi-dimensional matching and filtering were performed on the user's research interest profile and the matching category tags corresponding to multiple research projects to be classified, resulting in matching deviation values ​​for multiple research projects to be classified, including: Personalized feature vectors and domain preference weight tables are embedded into a unified feature space; within the unified feature space, the domain preference weight vectors corresponding to the matching classification labels of the research projects to be classified are determined. Within a unified feature space, the distance between the personalized feature vector and the domain preference weight vector, and the sum of the distances between the vector corresponding to the domain preference weight table and the domain preference weight vector are determined as the matching deviation values ​​for multiple research projects to be classified.

3. The method according to claim 2, characterized in that, The method further includes: Acquire users' real-time scientific research interaction data; determine multiple peer collaboration features based on personalized feature vectors, domain preference weight tables, and real-time scientific research interaction data through rule-based neighbor routing units; among which, real-time scientific research interaction data includes browsing projects, collecting literature, and peer collaboration records; Multiple peer collaboration features and personalized feature vectors are concatenated to obtain a fusion behavior feature vector; incremental data of user research interactions is obtained; the fusion behavior feature vector and incremental data of research interactions are iteratively updated through sliding window statistics to obtain an updated fusion behavior feature vector; among them, peer collaboration features and personalized feature vectors are concatenated with a weight of 7:

3. The updated fusion behavior feature vector and the research project attribute data of the research projects to be classified are mapped to the same feature space and standardized. Through cosine similarity matching, the user-project matching values ​​corresponding to multiple updated fusion behavior feature vectors and research projects to be classified are determined. Multiple target user-project matching values ​​that are greater than or equal to the preset user-project matching values ​​are determined. The target updated fusion behavior feature vectors corresponding to the multiple target user-project matching values ​​are used as research interest profiles.

4. The method according to claim 3, characterized in that, Through a rule-based neighbor routing unit, based on personalized feature vectors, domain preference weight tables, and real-time scientific research interaction data, multiple peer collaboration characteristics are determined, including: The user's core research field is determined by a domain preference weight table; multiple target real-time scientific research interaction data that match the core research field, meet the collaboration frequency standard, and have effective collaboration type are determined from real-time scientific research interaction data through preset rules. Based on the collaboration frequency corresponding to real-time scientific research interaction data of multiple targets, the collaboration weight is determined; the domain preference weight in the domain preference weight table is multiplied by the collaboration weight to obtain the modified domain preference weight table, and the modified domain preference weight table is used as the collaboration feature of multiple peers.

5. The method according to claim 4, characterized in that, Identify multiple target user-project matching values ​​that are greater than or equal to a preset user-project matching value, including: By leveraging the platform's user research database and using each user's unique identifier, historical classification feedback data is obtained from the user; and project classification screening thresholds are determined based on this historical classification feedback data. Based on the research stage label, determine the research stage adjustment coefficient; determine the product of the project classification screening threshold and the research stage adjustment coefficient as the stage project classification screening threshold; use the stage project classification screening threshold as the preset user-project matching value, and determine multiple target user-project matching values ​​that are greater than or equal to the preset user-project matching value.

6. The method according to claim 5, characterized in that, The domain popularity labeling unit is trained using the following method: The sample research interest profile, sample preference weight, sample real-time domain hot data, research project attribute data of multiple sample research projects to be classified, sample real-time labels of sample labeled research projects and sample popularity value are mapped to a unified vector space. The domain popularity labeling unit is trained by using sample research interest profiles, sample preference weights, sample real-time domain hot data, research project attribute data of multiple samples to be classified, sample real-time labels of multiple samples labeled research projects, and sample popularity values. The domain popularity labeling unit comprises an input layer, a feature alignment layer, a dual-branch matching layer, and a fusion calculation layer, arranged sequentially. The dual-branch matching layer includes a user interest matching branch layer and a domain hotspot association branch layer. The user interest matching branch layer is used to determine the basic matching degree between the research project attribute vector corresponding to the research project attribute data and the research interest vector corresponding to the research interest profile, and to determine the product of the basic matching degree and the preference weight vector corresponding to the preference weight, which serves as the weighted interest matching degree. The domain hotspot association branch layer is used to determine the basic correlation degree between the research project attribute vector and the real-time domain hotspot vector corresponding to the real-time domain hotspot data, and inputs the basic correlation degree and timeliness coefficient into the time decay weighting module, which outputs the corrected hotspot correlation degree. The weighted interest matching degree, the corrected hotspot correlation degree, the real-time labels of multiple sample labeled research projects, and the vectors corresponding to the sample popularity values ​​are input into the fusion calculation layer.

7. The method according to claim 1, characterized in that, The matching deviation values ​​of the user's current research needs and multiple unclassified research projects are used for requirement verification and optimization. This results in a requirement adaptation category set that adapts multiple unclassified research projects to the current research needs. The requirement adaptation category set includes multiple requirement adaptation categories that adapt to the current research needs, including: Obtain academic relevance weights, secondary field popularity weights, and secondary matching deviation weights; when the current research need is literature review, obtain the academic relevance indicators and secondary field popularity corresponding to the current research need. The product of academic relevance indicators and academic relevance weights, and the sum of the product of the second field popularity and the second field popularity weights, are determined as the academic relevance evaluation indicator; the product of the matching deviation value and the second matching deviation weights is determined as the matching deviation evaluation indicator; the difference between the academic relevance evaluation indicator and the matching deviation evaluation indicator is determined as the comprehensive evaluation indicator for multiple research projects to be classified; multiple research projects to be classified whose comprehensive evaluation indicator is greater than or equal to the preset comprehensive evaluation indicator are identified, thus obtaining a set of demand-fit classifications for multiple research projects to be classified to meet current research needs.

8. The method according to claim 7, characterized in that, The weight for feasibility is 0.3, the weight for popularity in the first field is 0.2, and the weight for the first matching deviation is 0.5; the weight for academic relevance is 0.4, the weight for popularity in the second field is 0.1, and the weight for the second matching deviation is 0.

5.

9. A scientific research project classification device based on personalized machine learning, characterized in that, Includes units for implementing the method of any one of claims 1 to 8.

Images