Self-adaptive multi-dimensional evaluation high-quality scientific and technological information screening method and system

Abstract

The invention discloses a self-adaptive multi-dimensional evaluation high-quality scientific and technological information screening method and system, and relates to the technical field of information screening, and the method comprises the following steps: collecting data from a database by using an API interface and a web crawler technology, and establishing a multi-source data set; establishing a preprocessing data set; performing feature extraction on the preprocessed data set, and establishing a multi-dimensional feature set after standardization processing; carrying out dimension evaluation under a six-dimensional evaluation channel on the multi-dimensional feature set, and establishing a dimension score; and after the dynamic weight mapped by the dimension score is adaptively configured, weighted calculation is executed, and a scientific and technological information screening result is generated. The technical problems that in the prior art, due to the fact that the scientific and technological information screening dimension is single, the evaluation weight is fixed and different scenes are difficult to adapt, the screening result is insufficient in accuracy and comprehensiveness are solved, and the purposes of achieving self-adaptive multi-dimensional evaluation and high-quality screening of the scientific and technological information and improving the screening efficiency are achieved. And the accuracy and comprehensiveness of scientific and technological information screening are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information screening, and in particular to a high-quality scientific and technological information screening method and system based on adaptive multi-dimensional evaluation. BACKGROUND

[0002] Under the background of current explosive growth of scientific and technological information, various databases, academic platforms and network resources are filled with massive scientific and technological literature, patents, achievements and other information. How to accurately screen out high-quality and high-value scientific and technological information from them has become the core demand of researchers, decision makers and related institutions. However, the existing technology mainly depends on single dimension (such as cited number, journal impact factor) or fixed weight evaluation system, which is difficult to cover the multi-aspect characteristics of scientific and technological information, and cannot dynamically adjust the evaluation standard according to different application scenarios (such as front technology tracking, achievement transformation evaluation), resulting in problems such as one-sidedness and poor adaptability of the screening results, which are difficult to meet the actual demand.

[0003] The existing technology has the technical problems of single dimension of scientific and technological information screening, fixed evaluation weight and difficulty in adapting to different scenarios, resulting in insufficient accuracy and lack of comprehensiveness of the screening results. SUMMARY

[0004] The present application provides a high-quality scientific and technological information screening method and system based on adaptive multi-dimensional evaluation, which is used to solve the technical problems of single dimension of scientific and technological information screening, fixed evaluation weight and difficulty in adapting to different scenarios in the prior art, resulting in insufficient accuracy and lack of comprehensiveness of the screening results.

[0005] In view of the above problems, the present application provides a high-quality scientific and technological information screening method and system based on adaptive multi-dimensional evaluation.

[0006] In a first aspect, the present application provides a high-quality scientific and technological information screening method based on adaptive multi-dimensional evaluation, which comprises:

[0007] Data is collected from the database by using API interface and web crawler technology to establish a multi-source data set; after data preprocessing of the multi-source data set, a pretreated data set is established, the data preprocessing includes HTML tag cleaning, format unification processing, stop word filtering processing, stem extraction, entity recognition and named entity normalization processing; the pretreated data set is subjected to feature extraction, the feature extraction includes text feature extraction, network feature calculation and time sequence feature construction, and a multi-dimensional feature set after standardization processing is established; the multi-dimensional feature set is subjected to dimension evaluation under six-dimensional evaluation channels to establish dimension scores; after adaptive configuration of dynamic weight of dimension score mapping, weighted calculation is performed to generate scientific and technological information screening results.

[0008] In a second aspect of the present application, a high-quality scientific and technological information screening system based on adaptive multi-dimensional evaluation is provided, and the system comprises:

[0009] A multi-source data set establishing module is configured to collect data from a database by using an API interface and a web crawler technology to establish a multi-source data set; a pre-processed data set establishing module is configured to establish a pre-processed data set after data preprocessing of the multi-source data set, wherein the data preprocessing comprises HTML tag cleaning, format unification processing, stop word filtering processing, stem extraction, entity recognition and named entity normalization processing; a multi-dimensional feature set establishing module is configured to establish a standardized multi-dimensional feature set after feature extraction of the pre-processed data set, wherein the feature extraction comprises text feature extraction, network feature calculation and time sequence feature construction; a dimension score establishing module is configured to establish a dimension score after dimension evaluation of the multi-dimensional feature set under a six-dimensional evaluation channel; and a screening result generating module is configured to generate a scientific and technological information screening result after adaptive configuration of dynamic weights of the dimension score mapping and execution of weighted calculation.

[0010] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0011] Data is collected from a database by using an API interface and a web crawler technology to establish a multi-source data set; a pre-processed data set is established after data preprocessing of the multi-source data set, wherein the data preprocessing comprises HTML tag cleaning, format unification processing, stop word filtering processing, stem extraction, entity recognition and named entity normalization processing; a standardized multi-dimensional feature set is established after feature extraction of the pre-processed data set, wherein the feature extraction comprises text feature extraction, network feature calculation and time sequence feature construction; a dimension score is established after dimension evaluation of the multi-dimensional feature set under a six-dimensional evaluation channel; and a scientific and technological information screening result is generated after adaptive configuration of dynamic weights of the dimension score mapping and execution of weighted calculation. The technical effect of realizing adaptive multi-dimensional evaluation and high-quality screening of scientific and technological information is achieved, and the accuracy and comprehensiveness of scientific and technological information screening are improved. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1 A high-quality scientific and technological information screening method flowchart based on adaptive multi-dimensional evaluation is provided for the embodiments of the present application.

[0014] Figure 2 This is a schematic diagram of the structure of a high-quality scientific and technological information screening system with adaptive multidimensional evaluation, provided in an embodiment of this application.

[0015] Figure labeling: Multi-source dataset creation module 10, preprocessing dataset creation module 20, multidimensional feature set creation module 30, dimensional score creation module 40, and filtering result generation module 50. Detailed Implementation

[0016] This application provides a high-quality scientific and technological information screening method and system with adaptive multidimensional evaluation, which is used to address the technical problems in the prior art where the screening dimensions of scientific and technological information are single, the evaluation weights are fixed and difficult to adapt to different scenarios, resulting in insufficient accuracy and comprehensiveness of the screening results.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, this application provides a method for screening high-quality scientific and technological information using adaptive multidimensional evaluation, the method comprising:

[0019] Step S100: Use API interfaces and web crawling technology to collect data from the database and establish a multi-source dataset.

[0020] Specifically, by calling the API interfaces of academic databases (such as Web of Science, Scopus, CNKI, etc.) and using web crawling technology to collect data from multiple channels such as patent databases and science and technology information platforms, the collected data covers various types such as academic papers, patent documents, and science and technology information. By integrating and aggregating this information from different data sources, a multi-source dataset is established, providing a comprehensive raw data foundation for subsequent information filtering and processing.

[0021] Step S200: After preprocessing the multi-source dataset, a preprocessed dataset is established. The data preprocessing includes HTML tag cleaning, format unification processing, stop word filtering processing, stemming extraction, entity recognition, and named entity normalization processing.

[0022] Specifically, the multi-source data set collected through the API interface and web crawler technology is pre-processed. First, the HTML tags contained in the data are cleaned to remove format interference, and then the data from different sources is uniformly processed to ensure data structure consistency. Then, stop word filtering is performed to remove common words without actual meaning, and the words are unified into stem form through stem extraction to reduce word form differences. At the same time, entity recognition technology is used to extract key entities in the data, and the named entities are normalized to eliminate different forms of the same entity. Finally, a standardized pre-processed data set is formed, laying a foundation for subsequent feature extraction.

[0023] Step S300: Feature extraction is performed on the pre-processed data set, including text feature extraction, network feature calculation, and time sequence feature construction to establish a multi-dimensional feature set after standardization.

[0024] Specifically, feature extraction is performed on the pre-processed data set. Text feature extraction calculates the term frequency-inverse document frequency matrix through TF-IDF vectorization processing of the pre-processed data set, generates semantic vector representation of each word using Word2Vec or BERT model, extracts 1-3-gram syntax features, and extracts latent topic distribution based on LDA model. Network feature calculation performs citation network analysis by calculating node centrality indicators, constructs scholar influence map to analyze scholar collaboration and influence, constructs institution association network and evaluates the cooperation strength between institutions based on it. Time sequence feature construction identifies hot and cold spots in research fields by analyzing the publication time of literature, evaluates time sequence academic influence features according to the citation growth trend of literature, and uses the features for time decay analysis to quantify the timeliness features of literature. After standardizing the above extracted features, a multi-dimensional feature set is formed.

[0025] Step S400: Dimension evaluation under six-dimensional evaluation channel is performed on the multi-dimensional feature set to establish dimension scores.

[0026] Specifically, the standardized multi-dimensional feature set activates a six-dimensional evaluation channel to conduct dimension evaluation to establish a dimension score. First, an accuracy evaluation layer is activated, data evaluation is conducted based on citation verification rate, peer review status and data source reliability assessment, and a first dimension score is obtained. An timeliness evaluation layer is activated, data evaluation is conducted according to publication time decay function, update frequency and real-time index, and a second dimension score is generated. An authority evaluation layer is activated, data evaluation is conducted through author H-index, journal impact factor, institution ranking and academic reputation, and a third dimension score is determined. An importance evaluation layer is activated, data evaluation is conducted in combination with cited frequency, download volume, social media mention degree and academic attention degree, and a fourth dimension score is obtained. A field relevance evaluation layer is activated, data evaluation is conducted based on TF-IDF (term frequency-inverse document frequency) score, semantic similarity and topic model matching degree, and a fifth dimension score is obtained. A compliance evaluation layer is activated, data evaluation is conducted according to ethical approval status, regulation compliance check, data privacy protection and academic integrity, and a sixth dimension score is generated. Finally, the first to sixth dimension scores are integrated to form the dimension score.

[0027] Step S500: After adaptively configuring the dynamic weight of the dimension score mapping, weighted calculation is performed to generate a science and technology information screening result.

[0028] Specifically, the dimension score indicators obtained by the six-dimensional evaluation channel are first normalized to eliminate dimensional differences and obtain a normalized result. Then, the proportion of each indicator in all schemes in the normalized result is calculated by entropy weight method. The proportion value of each indicator is obtained by weighted sum of the normalized result, and the entropy value representing information uncertainty is calculated using the proportion value. Then, the indicator difference coefficient is obtained according to the entropy value, to establish the first weight result (objective weight). At the same time, an expert judgment matrix is constructed, and after passing the consistency check, the subjective weight of each indicator is calculated from the matrix using the eigenvector method as the second weight result. Then, the first weight result and the second weight result are fused, and the objective and subjective weights are dynamically balanced by the adaptive fusion coefficient to establish the dynamic weight of the dimension score mapping. Finally, the weighted calculation is performed on each dimension score based on the dynamic weight, and the comprehensive score of the science and technology information is obtained, and the science and technology information screening result is generated accordingly.

[0029] In one possible implementation manner, step S400 further includes:

[0030] Step S410: The accuracy evaluation layer in the six-dimensional evaluation channel is activated, and data evaluation based on citation verification rate, peer review status and data source reliability assessment is conducted using the accuracy evaluation layer to establish a first dimension score.

[0031] Step S420: activate the timeliness evaluation layer in the six-dimensional evaluation channel, use the timeliness evaluation layer to carry out data evaluation based on the publication time decay function, update frequency and real-time index, and establish the second-dimensional score.

[0032] Step S430: activate the authority evaluation layer in the six-dimensional evaluation channel, use the authority evaluation layer to carry out data evaluation based on the author H index, journal impact factor, institution ranking and academic reputation, and establish the third-dimensional score.

[0033] Step S440: establish the dimensional score according to the first-dimensional score, the second-dimensional score and the third-dimensional score.

[0034] Specifically, after activating the accuracy evaluation layer in the six-dimensional evaluation channel, the data evaluation is carried out in the following way to establish the first-dimensional score: the citation verification rate is calculated, that is, the ratio of the number of effective citations to the total number of citations; the peer review status is evaluated, and the corresponding score is obtained by combining the established review system of the journal and the evaluation of the reviewers; the reliability of the data source is quantitatively evaluated by using a machine learning classifier to determine the data source reliability; finally, the citation verification rate, the peer review score and the data source reliability are weighted and summed according to the preset weights (such as w1, w2, w3, and w1+w2+w3=1) to obtain the first-dimensional score.

[0035] After activating the timeliness evaluation layer in the six-dimensional evaluation channel, the data evaluation is carried out in the following way to establish the second-dimensional score: an exponential decay model is used to construct a publication time decay function, the formula is exp(-λ×(current time-publication time)), where λ is a decay coefficient dynamically adjusted according to the characteristics of the field; the update frequency of the data is counted and converted into a corresponding coefficient; the real-time index reflecting the current attention degree of the information is calculated; finally, the publication time decay function result, the update frequency coefficient and the real-time index are fused and calculated according to the preset weight to obtain the second-dimensional score, so as to quantify the timeliness value of the scientific and technological information.

[0036] After activating the authority evaluation layer in the six-dimensional evaluation channel, the H index of the author is collected first, and the author authority score is obtained by calculating Σ(H index i×weight i), where the weight i is set according to the contribution proportion of the author in the research; for the journal authority, the SCI / EI impact factor data and the reputation ranking in the industry of the journal are integrated, and the journal authority score is formed after standardization processing; for the institution authority, the institution ranking data in the authoritative ranking system such as QS and ESI are cited, which are quantified into corresponding scores; at the same time, the general recognition degree of the academic community to the author, the journal and the institution is combined to give the academic reputation a corresponding score; then an academic knowledge graph covering the author, the journal and the institution is constructed, and the authority propagation effect of each subject is calculated by using PageRank and other graph algorithms, and the four scores are weighted and fused to finally obtain the third-dimensional score.

[0037] After obtaining the first dimension score (the accuracy evaluation layer is based on the citation verification rate, peer review status and data source reliability evaluation), the second dimension score (the timeliness evaluation layer is based on the publication time decay function, update frequency and real-time index), and the third dimension score (the authority evaluation layer is based on the author H index, journal impact factor, institution ranking and academic reputation), the scores of the three dimensions are first normalized to eliminate dimensional differences, and then mapped to the [0, 1] interval. According to the basic importance of each dimension in the screening of scientific and technological information, the initial weights are preset (such as accuracy weight 0.4, timeliness weight 0.3, authority weight 0.3), and the integrated results are calculated by the weighted sum formula (dimension score = first dimension score x 0.4 + second dimension score x 0.3 + third dimension score x 0.3) to form the preliminary dimension score, which lays the foundation for subsequent fusion with other dimension scores and dynamic weight configuration.

[0038] In one possible implementation manner, step S440 further includes:

[0039] Step S441: activating the importance evaluation layer in the six-dimensional evaluation channel, using the importance evaluation layer to perform data evaluation based on cited frequency, download volume, social media mention degree, and academic attention degree, and establishing a fourth dimension score.

[0040] Step S442: activating the field relevance evaluation layer in the six-dimensional evaluation channel, using the field relevance evaluation layer to perform data evaluation based on TF-IDF score, semantic similarity, and topic model matching degree, and establishing a fifth dimension score.

[0041] Step S443: activating the compliance evaluation layer in the six-dimensional evaluation channel, using the compliance evaluation layer to perform data evaluation based on ethical approval status, regulation compliance check, data privacy protection, and academic integrity, and establishing a sixth dimension score.

[0042] Step S444: establishing a dimension score according to the first dimension score, the second dimension score, the third dimension score, the fourth dimension score, the fifth dimension score, and the sixth dimension score.

[0043] Specifically, after activating the importance evaluation layer in the six-dimensional evaluation channel, the citation frequency, download volume, social media mention degree, and academic attention degree are collected and normalized, and each index is uniformly mapped to the interval [0, 1]. The citation frequency and download volume are standardized by counting the original data and the maximum value in the field, the social media mention degree is quantified by combining the platform heat algorithm, and the academic attention degree is assigned by analyzing the research trend in the field. Then, the weights of each index (a, b, g, etc., and the sum of the weights is 1) are determined by multi-source data fusion technology, and the fourth-dimensional score is integrated by weighted calculation of a x normalized citation frequency + b x normalized download volume + g x social influence index, to quantify the importance of scientific and technological information.

[0044] After activating the field relevance evaluation layer in the six-dimensional evaluation channel, first, the TF-IDF (Term Frequency-Inverse Document Frequency) score calculation is carried out on the preprocessed data set. By performing word segmentation on the text, the term frequency (TFi) of each word in the current text and the inverse document frequency (IDFi) of the word in all documents are counted. Then, the cosine similarity between the text and the target domain keyword vector is combined to obtain the TF-IDF score through the formula Σ(TFi x IDFi x cosine-similarity), which quantifies the matching degree at the keyword level. Second, semantic similarity calculation is carried out. The Word2Vec model is used to train the words in the preprocessed data set to generate semantic vectors, or the BERT model is used to encode the text to obtain context-related semantic vectors. Then, the cosine distance between the text vector and the target domain benchmark vector is calculated to obtain the semantic similarity score. Third, theme model matching degree analysis is carried out. Based on the LDA model, the preprocessed data set is trained to extract the latent theme distribution of the text, and the standard theme distribution of the target domain is obtained. By calculating the KL divergence or JS distance between the two distributions, the theme matching degree score is obtained. Finally, the above TF-IDF score, semantic similarity score, and theme model matching degree score are normalized to be in the same numerical interval, and then weighted summation is carried out according to the preset weights (such as 0.3, 0.4, and 0.3 for each weight), to generate the fifth-dimensional score, which comprehensively quantifies the relevance of scientific and technological information and the target domain.

[0045] After the compliance evaluation layer in the six-dimensional evaluation channel is activated, data evaluation is performed in the following manner to establish the sixth-dimensional score: for the ethical approval status, it is checked whether the research involved in the scientific and technological information has passed the ethical committee approval and obtained the corresponding proof documents, and the approval level (such as passing completely, passing with conditions, and failing) is quantified as an ethical score; a regulation compliance check is carried out, and according to relevant field laws and regulations (such as academic research specifications, data use regulations, etc.), the compliance points in the information are automatically matched through a rule engine to identify potential violations and convert them into a regulation score; the data privacy protection situation is evaluated, and it is checked whether the information contains sensitive personal data, whether encryption or anonymization processing is used, etc., and a privacy score is generated through a privacy protection evaluation model; academic integrity is investigated, and a text duplication detection tool is used to check plagiarism, data falsification, etc., and the academic integrity score is determined according to the results of the investigation and the reports of the same industry; and finally, the ethical score, the regulation score, and the privacy score are weighted and summed according to the preset compliance evaluation weights (such as the ethical compliance weight, the regulation compliance weight, and the privacy protection weight, and the sum of the weights is 1) to obtain the sixth-dimensional score, so as to quantify the compliance level of the scientific and technological information.

[0046] After the first to sixth-dimensional scores are obtained, the dimensional scores are first normalized to map them to the [0, 1] interval to eliminate dimensional differences, wherein the positive indicators use the formula y = (x-min) / (max-min), and the negative indicators use the formula y = (max-x) / (max-min); then, based on the design logic of the six-dimensional evaluation channel, the initial weights of each dimension in the comprehensive evaluation of scientific and technological information are preset (such as accuracy 0.2, timeliness 0.15, authority 0.2, importance 0.15, field relevance 0.15, and compliance 0.15, and the sum of the weights is 1); and then the normalized six-dimensional scores are integrated and calculated through the weighted sum formula (dimensional score = first-dimensional score*0.2 + second-dimensional score*0.15 + third-dimensional score*0.2 + fourth-dimensional score*0.15 + fifth-dimensional score*0.15 + sixth-dimensional score*0.15) to finally obtain the dimensional score that can fully reflect the comprehensive quality of the scientific and technological information, providing basic data for subsequent adaptive weight configuration and screening result generation.

[0047] In one possible implementation manner, step S500 further includes:

[0048] Step S510: Normalizing the indicators of different dimensional scores to establish the normalization result of the calibration interval.

[0049] Step S520: Calculating the proportion of the indicators in all schemes according to the entropy weight method to establish the first weight result.

[0050] Step S530: Obtaining the subjective weight of the indicators to establish the second weight result.

[0051] Step S540: After the first weight result and the second weight result are fused, a dynamic weight is established.

[0052] Specifically, for each index involved in the first to sixth dimension scores (such as the citation verification rate, the publication time decay value, the author H-index, the cited frequency, the TF-IDF score, the ethical approval score, etc.), the positive index and the negative index are distinguished for normalization processing. Among them, the positive index (the larger the numerical value, the better the quality, such as the citation verification rate, the cited frequency, etc.) is calculated by the formula y = (x-min) / (max-min), and the negative index (the smaller the numerical value, the better the quality, such as the institution ranking, etc.) is calculated by the formula y = (max-x) / (max-min). All index values are uniformly converted into the calibrated interval of [0, 1] to eliminate the dimensional difference and numerical range difference between different indexes, forming a standardized normalized result to provide comparable basic data for subsequent weight calculation.

[0053] Based on the obtained normalized result, the entropy weight method is used to calculate the proportion of each index in all schemes to establish the first weight result. First, for each index, the proportion value p ij of each index in different schemes is calculated, and the formula is p ij = y ij / ∑y ij (where y ij is the normalized value of the jth index in the ith scheme); then, the entropy value Ej of each index is calculated according to the proportion value, and the formula is Ej = -k x ∑(p ij x ln(p ij )), where k is a constant and k = 1 / ln(n), and n is the total number of schemes; subsequently, the difference coefficient dj of each index is calculated through the entropy value, that is, dj = 1-Ej; finally, the difference coefficient is normalized to obtain the first weight result (objective weight), and the formula is wj = dj / ∑dj, so as to reflect the objective importance of each index based on the information of the data itself.

[0054] The subjective weight of the indexes is obtained by constructing an expert judgment matrix to establish the second weight result. First, senior experts in the field are invited to compare and assign values (for example, using the 1-9 scale method, 1 representing equal importance and 9 representing extreme importance) to the indexes according to their relative importance in the screening of scientific and technological information, forming an n*n judgment matrix (n is the number of indexes); then, the consistency of the judgment matrix is checked, the consistency index CI = (λmax-n) / (n-1) is calculated (where λmax is the maximum eigenvalue of the matrix), and the consistency ratio CR = CI / RI is determined in combination with the average random consistency index RI. When CR < 0.1, the matrix passes the consistency test; after the test passes, the characteristic vector method is used to solve the maximum eigenvalue corresponding to the characteristic vector of the judgment matrix, and the characteristic vector is normalized to obtain the subjective weight of each index, that is, the second weight result, to reflect the subjective judgment of the importance of the index by experts' experience.

[0055] The first weight result (objective weight) and the second weight result (subjective weight) obtained are fused to establish a dynamic weight. First, an adaptive fusion coefficient a is introduced, which is dynamically adjusted according to factors such as data dispersion, user actual demand preference, and historical screening effect feedback (a takes a value in the range [0, 1]); then, the weighted formula wj = a x wj obj +(1-a) x wj sub The two weights are fused, where wj obj is the weight of the jth index in the first weight result, and wj sub is the weight of the jth index in the second weight result; the wj formed after fusion is the dynamic weight, which can flexibly balance the influence of objective data information and subjective expert experience according to the actual scene, realizing dynamic adaptation of the weight.

[0056] In one possible implementation, step S520 further includes:

[0057] Step S521: Weighted sum of the normalized results to calculate the proportion value of each index.

[0058] Step S522: Calculate the entropy value of each index using the proportion value, which represents the information uncertainty of the index.

[0059] Step S523: Calculate the index difference coefficient according to the entropy value, and use the index difference coefficient to establish the first weight result.

[0060] Specifically, for the obtained normalized results, all schemes and corresponding indexes are traversed, and the normalized values of each index in different schemes are accumulated and summed to obtain the total normalized value of the index; then, the normalized value of the index in each scheme is compared with the total normalized value, that is, the formula p ij=y ij / Σy ij (wherein y ij is the normalized value of the jth indicator in the ith scheme, and Σy ij is the sum of the normalized values of the jth indicator in all schemes), the proportion value p ij of each indicator in each scheme is calculated to quantify the relative proportion of the indicator in a specific scheme.

[0061] Based on the proportion value p ij of each indicator in different schemes, the entropy value of each indicator is calculated by the entropy calculation formula Ej=-k×Σ(p ij ×ln(p ij )), where k is a constant and k=1 / ln(n) (n is the total number of schemes), and when p ij =0, p ij ×ln(p ij ) is processed as 0; the greater the calculated entropy value Ej, the more uniform the proportion distribution of the indicator in different schemes, the less the effective distinguishing information contained, and the higher the information uncertainty; on the contrary, the smaller the entropy value, the more significant the difference of the indicator between schemes, the lower the information uncertainty, and the effective distinguishing ability of different schemes.

[0062] Based on the entropy value Ej of each indicator, the difference coefficient dj of each indicator is calculated by the formula dj=1-Ej, and the size of the difference coefficient reflects the distinguishing ability of the indicator between different schemes, and the greater dj indicates that the indicator has more significant distinguishing effect on the scheme evaluation; subsequently, the difference coefficients of all indicators are normalized, that is, the weight of each indicator is calculated by the formula wj=dj / Σdj (where Σdj is the sum of the difference coefficients of all indicators), and thus the first weight result (objective weight) is established, which objectively reflects the importance of each indicator based on data information, and the indicator with a greater difference coefficient has a higher proportion in the first weight result.

[0063] In one possible implementation manner, the step S530 further includes:

[0064] Step S531: constructing an expert judgment matrix according to the importance of the indicator evaluation factors.

[0065] Step S532: performing consistency checking of the matrix on the expert judgment matrix.

[0066] Step S533: if the consistency checking is passed, calculating the weight of each indicator from the expert judgment matrix by the eigenvector method to establish the subjective weight.

[0067] Specifically, according to the importance of each indicator in the evaluation system, experts with rich experience in the field are invited to compare all indicators pairwise and assign corresponding values. A 1-9 scale method is used to quantify the importance difference, where 1 means that the importance of the two indicators is the same; 3 means that the importance of the first indicator is slightly greater than that of the second indicator; 5 means that the importance of the first indicator is obviously greater than that of the second indicator; 7 means that the importance of the first indicator is strongly greater than that of the second indicator; 9 means that the importance of the first indicator is extremely greater than that of the second indicator; if the importance of the second indicator is greater than that of the first indicator, the reciprocal of the corresponding value is taken. In this way, an n x n expert judgment matrix A = [a ij ](n is the number of indicators) is constructed, where the elements a ij in the matrix represent the importance ratio of the ith indicator to the jth indicator.

[0068] For the constructed expert judgment matrix, the consistency of the matrix is checked to ensure the rationality of the judgment logic. First, the maximum eigenvalue λmax of the matrix is calculated, and the consistency index CI is calculated by the formula CI = (λmax-n) / (n-1) (where n is the number of indicators); then, according to the number of indicators n, the corresponding RI value is obtained from the pre-set average random consistency index table; and then the consistency ratio CR is calculated by the formula CR = CI / RI, when CR < 0.1, it is determined that the expert judgment matrix passes the consistency check, indicating that the expert's judgment has logical consistency; if CR ≥ 0.1, the expert needs to adjust the judgment matrix again until the consistency requirement is met.

[0069] When the expert judgment matrix passes the consistency check, the characteristic vector method is used to calculate the weight of each indicator to establish the subjective weight. First, the characteristic equation |A-λI| = 0 (where A is the expert judgment matrix, λ is the eigenvalue, and I is the unit matrix) is solved to obtain the characteristic vector corresponding to the maximum eigenvalue λmax; then, the characteristic vector is normalized, that is, each element in the vector is divided by the sum of all elements, so that the sum of the elements in the processed vector is 1; the normalized characteristic vector is the subjective weight of each indicator, and the value of each element corresponds to the subjective importance of the corresponding indicator in the evaluation system, thereby completing the establishment of the subjective weight.

[0070] In one possible implementation, step S300 further includes:

[0071] Step S310: performing TF-IDF vectorization processing on the preprocessed data set to calculate a term frequency-inverse document frequency matrix.

[0072] Step S320: using a Word2Vec or BERT model to generate a semantic vector representation of each word in the preprocessed data set.

[0073] Step S330: 1-3-gram syntax feature extraction is performed on the preprocessed data set.

[0074] Step S340: latent topic distribution of the preprocessed data set is extracted based on the LDA model.

[0075] Specifically, for the preprocessed data set (text data that has completed cleaning, word segmentation, stop word removal, etc.), TF-IDF vectorization processing is performed to calculate the term frequency-inverse document frequency matrix. First, the number of occurrences of each word in a single document is counted to obtain the term frequency (TF), i.e., TF = number of occurrences of a word in a document / total number of words in the document; second, the inverse document frequency (IDF) is calculated, with the formula being IDF = log(total number of documents / (number of documents containing the word + 1)), and the +1 is to avoid a denominator of 0; then, the TF and IDF of each word are multiplied to obtain the TF-IDF value of the word in the corresponding document; finally, the term frequency-inverse document frequency matrix is constructed with the document as the row, the word as the column, and the TF-IDF value as the element, and each element in the matrix directly reflects the importance weight of the corresponding word to a specific document.

[0076] For the preprocessed data set (text data that has been cleaned, segmented, etc.), when using the Word2Vec model, the text sequence is traversed by setting a sliding window, the context words within the window are taken as input with the target word as the center, and the model is trained through the Skip-gram or CBOW architecture to learn the semantic association between words, and finally a fixed-dimensional (such as 100-dimensional or 200-dimensional) word vector is output, each vector corresponding to a word in the data set representing its semantic information; if the BERT model is used, the preprocessed text is segmented according to the model requirements (such as WordPiece segmentation) and special markers ([CLS], [SEP]) are added, and then input into the pre-trained BERT model, and the self-attention mechanism of the multi-layer Transformer encoder is used to capture the context dependency relationship, and the dynamic vector representation of each word in a specific context is extracted, thereby generating a word vector that can reflect the context semantics, and realizing the semantic vector representation of each word in the preprocessed data set.

[0077] For the pre-processed data set (text data that has been cleaned, segmented, etc.), a sliding window method is used to extract 1-3-gram syntax features. For 1-gram, each independent word in the data set is directly extracted as a feature to capture the information of individual words. For 2-gram, two adjacent words are taken as a unit, and the text is traversed through a sliding window to extract word pairs (such as "technology and innovation" and "data and analysis") composed of two consecutive words as features. For 3-gram, three consecutive words are taken as a unit, and the combination of words (such as "artificial intelligence, application, and scenario" and "pre-processing, data, and set") is extracted through a sliding window as a feature. In this way, the features of word sequences of different lengths in the text are comprehensively captured, the local word order and word combination information are retained, and rich syntax level feature support is provided for subsequent text analysis and model training.

[0078] For the pre-processed data set (text data that has been cleaned, segmented, and stop words removed), the latent topic distribution is extracted based on the LDA (Latent Dirichlet Allocation) model. First, a reasonable number of topics K is preset, each document is treated as a mixed distribution of multiple topics, and each topic is treated as a probability distribution of multiple words. Then, the model is trained by methods such as Gibbs sampling or variational inference to iteratively optimize the parameters of the document-topic distribution and the topic-word distribution. During the training process, the probability of each word belonging to a specific topic and the probability of each document containing a specific topic are constantly adjusted until the model converges. Finally, the probability distribution of each document on K topics (i.e., the document-topic distribution) and the high-frequency words contained in each topic and their probabilities (i.e., the topic-word distribution) in the pre-processed data set are output, clearly presenting the topic structure and core content of each topic in the data set, and completing the extraction of the latent topic distribution.

[0079] In one possible implementation, step S300 further includes:

[0080] Step S350: performing citation network analysis by calculating node centrality indicators.

[0081] Step S360: constructing a scholar influence map and performing scholar collaboration relationship and influence analysis.

[0082] Step S370: constructing an institution correlation network and evaluating the collaboration strength between structures based on the institution correlation network.

[0083] Specifically, for the constructed citation network (in which nodes represent literatures and edges represent citation relationships between literatures), citation network analysis is performed by calculating node centrality indicators. Degree centrality is calculated to measure the number of times a literature is directly cited by other literatures, and the more times, the stronger the direct influence in the network; betweenness centrality is calculated to evaluate the pivotal role of a literature in connecting different literature communities, and the higher the value, the more critical it is in transmitting citation relationships; closeness centrality is calculated to reflect the average distance of a literature to other literatures in the network, and the closer the distance, the faster it can be associated with other literatures. Through comprehensive analysis of these indicators, core literatures, key intermediate literatures and literatures with extensive connectivity in the citation network can be identified, thereby revealing the structural characteristics and information transmission paths of the network.

[0084] The academic data of scholars are collected, including published papers (including journal names, publication times, co-authors), cited data (total citation frequency, single highest citation, H index), etc., to establish a structured database; then taking scholars as nodes, based on the number and frequency of co-published papers, the network visualization tool Gephi is used to construct a scholar collaboration network map, and the degree centrality and closeness centrality of the nodes in the network are calculated to identify core collaboration nodes and collaboration communities; at the same time, the cited indicators are associated with the collaboration network data, and a weighted summation method (such as H index accounting for 40%, total citation frequency accounting for 30%, and collaboration network centrality accounting for 30%) is used to calculate the comprehensive influence score, and the size and color depth of the nodes in the map intuitively present the differences in the influence of scholars, and finally realize the quantitative and visual analysis of the collaboration mode (such as cross-institutional collaboration frequency, core collaboration circle) and influence level of scholars.

[0085] The collaboration information between institutions is collected, including the number of co-published papers, the number and level of jointly applied scientific research projects, and the citation of collaboration results, etc., to construct an institutional correlation network based on this. Each node in the network represents an institution, and the edges between nodes represent the existence of cooperation between institutions. The weight of the edge is determined comprehensively according to the cooperation frequency, cooperation result quality and other factors (such as co-publishing 1 core journal paper for 2 points, co-applying 1 national-level project for 5 points, and assigning weights according to the total score); then, based on the network, the cooperation strength between institutions is calculated, the cooperation tightness is directly judged by analyzing the weight value of the edge, and the density (the ratio of actual cooperation edges to possible cooperation edges) and clustering coefficient (the tightness of the institution collaboration circle) of the network are calculated to comprehensively evaluate the depth and breadth of cooperation between different institutions, and clarify the overall structure and key cooperation links of the institutional cooperation network.

[0086] In one possible implementation manner, the step S300 further includes:

[0087] By analyzing the publication time of the literature, the hot and cold points of the research field are identified.

[0088] According to the citation growth trend of the literature, the time sequence academic influence feature of the literature is evaluated.

[0089] The time decay analysis is performed using the time sequence academic influence feature, and the timeliness feature of the literature is quantified.

[0090] Specifically, by analyzing the publication time of each literature in the above-mentioned file, the number of literatures published in different time nodes (such as per year, per quarter) in a specific research field is counted, and the correspondence between time and the number of literatures is constructed. If the number of literatures published in a certain period of time shows a significant and continuous growth trend, even an explosive growth, and the trend remains stable or continues to rise in the subsequent period of time, it can be determined that the research direction corresponding to the period of time is a hot field; on the contrary, if the number of literatures published in a certain period of time is long-term at a low level, grows slowly or even shows a downward trend, and there is no obvious warming in a long period of time, the research direction corresponding to the period of time is identified as a cold field, so as to clarify the heat change of the research field in the time dimension.

[0091] The publication time and the number of citations of each literature in the above-mentioned file are extracted, and the citation growth curve is drawn with time as the horizontal axis and the number of citations as the vertical axis; by calculating the initial growth rate (the average annual increase of citations in the first 1-3 years after publication), the peak occurrence time (the year when the number of citations reaches the maximum), the peak duration (the years maintaining more than 80% of the peak value), and the decay rate (the average annual decrease of citations after the peak), etc. Parameters, the time sequence feature is quantified, if the initial growth rate is > 50 times / year, the peak duration is > 5 years, and the decay rate is < 10% per year, it is determined as “sustained high influence”; if the initial growth rate is > 100 times / year but the peak duration is < 2 years, it is determined as “short-term explosive influence”; if the initial growth rate is < 10 times / year and there is a slow upward trend (annual increase > 5%) after 10 years, it is determined as “gradual diffusion influence”, so as to realize the accurate evaluation of the time sequence academic influence feature of the literature.

[0092] The time decay analysis is performed using the time sequence academic influence feature (such as the peak value and decay rate of the citation growth curve) of the above-mentioned file, and the exponential decay model (formula C(t) = C0×e (-kt)Wherein C(t) is the number of citations after t years of publication of the literature, C0 is the citation base at the initial stage of publication, and k is the attenuation coefficient) to quantify the timeliness feature. The reference data of each literature is fitted with the model, and the attenuation coefficient k is calculated. The larger the value of k, the faster the number of citations of the literature decreases over time, the stronger the timeliness, and the more the academic value is reflected in the short term; the smaller the value of k, the slower the attenuation of the number of citations, the weaker the timeliness, and thus the specific value of k directly presents the timeliness difference of each literature.

[0093] Embodiment two, based on the same inventive concept as the high-quality scientific and technological information screening method of adaptive multi-dimensional evaluation in the foregoing embodiment, as Figure 2 The application provides a high-quality scientific and technological information screening system based on adaptive multi-dimensional evaluation, and the system and method embodiments in the application are based on the same inventive concept. The system comprises:

[0094] The multi-source data set establishment module 10 is configured to collect data from a database by using an API interface and a web crawler technology, and establish a multi-source data set.

[0095] The preprocessed data set establishment module 20 is configured to establish a preprocessed data set by performing data preprocessing on the multi-source data set, wherein the data preprocessing comprises HTML tag cleaning, format unification processing, stop word filtering processing, stem extraction, entity recognition and named entity normalization processing.

[0096] The multi-dimensional feature set establishment module 30 is configured to perform feature extraction on the preprocessed data set, wherein the feature extraction comprises text feature extraction, network feature calculation and time series feature construction, and a standardized multi-dimensional feature set is established.

[0097] The dimension score establishment module 40 is configured to perform dimension evaluation on the multi-dimensional feature set under six-dimensional evaluation channels, and establish a dimension score.

[0098] The screening result generation module 50 is configured to perform weighted calculation after adaptively configuring dynamic weights of the dimension score mapping, and generate a scientific and technological information screening result.

[0099] Further, the system is also configured to implement the following functions:

[0100] The accuracy evaluation layer in the six-dimensional evaluation channel is activated, the accuracy evaluation layer is used for data evaluation based on the citation verification rate, peer review status and data source reliability evaluation, and a first-dimensional score is established; the timeliness evaluation layer in the six-dimensional evaluation channel is activated, the timeliness evaluation layer is used for data evaluation based on the publication time decay function, update frequency and real-time index, and a second-dimensional score is established; the authority evaluation layer in the six-dimensional evaluation channel is activated, the authority evaluation layer is used for data evaluation based on the author H index, journal impact factor, institution ranking and academic reputation, and a third-dimensional score is established; and a dimensional score is established according to the first-dimensional score, the second-dimensional score and the third-dimensional score.

[0101] Further, the system is also used to implement the following functions:

[0102] The importance evaluation layer in the six-dimensional evaluation channel is activated, the importance evaluation layer is used for data evaluation based on the cited frequency, download volume, social media mention degree and academic attention degree, and a fourth-dimensional score is established; the field relevance evaluation layer in the six-dimensional evaluation channel is activated, the field relevance evaluation layer is used for data evaluation based on the TF-IDF score, semantic similarity and topic model matching degree, and a fifth-dimensional score is established; the compliance evaluation layer in the six-dimensional evaluation channel is activated, the compliance evaluation layer is used for data evaluation based on the ethical approval status, regulation compliance check, data privacy protection and academic integrity, and a sixth-dimensional score is established; and a dimensional score is established according to the first-dimensional score, the second-dimensional score, the third-dimensional score, the fourth-dimensional score, the fifth-dimensional score and the sixth-dimensional score.

[0103] Further, the system is also used to implement the following functions:

[0104] The indicators of different dimensional scores are normalized to establish a normalized result of a calibration interval; the normalized result is used to calculate the proportion of the indicators in all schemes according to the entropy weight method to establish a first weight result; the subjective weight of the indicators is obtained to establish a second weight result; and the first weight result and the second weight result are fused to establish a dynamic weight.

[0105] Further, the system is also used to implement the following functions:

[0106] The normalized results are weighted and summed to calculate the proportion value of each indicator; the proportion value is used to calculate the entropy value of each indicator, and the entropy value represents the information uncertainty of the indicator; and the indicator difference coefficient is calculated according to the entropy value, and the first weight result is established by using the indicator difference coefficient.

[0107] Further, the system is also used to implement the following functions:

[0108] According to the importance of the index evaluation factors, an expert judgment matrix is constructed; the consistency of the matrix is checked; if the consistency check is passed, the weight of each index is calculated from the expert judgment matrix by using the characteristic vector method, and the subjective weight is established.

[0109] Further, the system is also used to implement the following functions:

[0110] The pre-processed data set is subjected to TF-IDF vectorization processing to calculate the term frequency-inverse document frequency matrix; a semantic vector representation of each word in the pre-processed data set is generated using a Word2Vec or BERT model; 1-3 grammar features of the pre-processed data set are extracted; and the latent topic distribution of the pre-processed data set is extracted based on an LDA model.

[0111] Further, the system is also used to implement the following functions:

[0112] The reference network analysis is performed by calculating the node centrality index; the scholar influence map is constructed to perform the scholar cooperation relationship and influence analysis; the institution correlation network is constructed to evaluate the cooperation strength between structures based on the institution correlation network.

[0113] Further, the system is also used to implement the following functions:

[0114] The hot and cold spots of the research field are identified by analyzing the publication time of the literature; the time sequence academic influence feature of the literature is evaluated according to the citation growth trend of the literature; and the time decay analysis is performed using the time sequence academic influence feature to quantify the timeliness feature of the literature.

[0115] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above describes a specific embodiment of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0116] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0117] The present application is only an exemplary description of the present application, and should be considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalents, the present application is intended to include these modifications and changes.

Claims

1. A method for high quality technology information filtering by adaptive multidimensional evaluation, characterized in that, The method comprises: Collecting data from the database by using API interface and web crawler technology, and establishing a multi-source data set; After data preprocessing of the multi-source data set, a pretreated data set is established, and the data preprocessing includes HTML tag cleaning, format unification processing, stop word filtering processing, stem extraction, entity recognition and named entity normalization processing; After feature extraction of the pretreated data set, a multi-dimensional feature set after standardization processing is established, and the feature extraction includes text feature extraction, network feature calculation and time sequence feature construction; Dimension evaluation under a six-dimensional evaluation channel is performed on the multi-dimensional feature set, and a dimension score is established; After adaptively configuring dynamic weights of the dimension score mapping, weighted calculation is performed to generate a science and technology information screening result.

2. The method for screening high-quality scientific and technological information through self-adaptive multi-dimensional evaluation according to claim 1, characterized in that, The dimension evaluation under the six-dimensional evaluation channel and the establishment of the dimension score include: An accuracy evaluation layer in the six-dimensional evaluation channel is activated, data evaluation based on citation verification rate, peer review status and data source reliability evaluation is performed by using the accuracy evaluation layer, and a first dimension score is established; An timeliness evaluation layer in the six-dimensional evaluation channel is activated, data evaluation based on publication time decay function, update frequency and real-time index is performed by using the timeliness evaluation layer, and a second dimension score is established; An authority evaluation layer in the six-dimensional evaluation channel is activated, data evaluation based on author H index, journal impact factor, institution ranking and academic reputation is performed by using the authority evaluation layer, and a third dimension score is established; The dimension score is established according to the first dimension score, the second dimension score and the third dimension score.

3. A method for high quality technology information filtering with self-adaptive multi-dimensional evaluation as claimed in claim 2, characterized in that, The establishment of the dimension score according to the first dimension score, the second dimension score and the third dimension score includes: An importance evaluation layer in the six-dimensional evaluation channel is activated, data evaluation based on cited frequency, download volume, social media mention degree and academic attention degree is performed by using the importance evaluation layer, and a fourth dimension score is established; A field relevance evaluation layer in the six-dimensional evaluation channel is activated, data evaluation based on TF-IDF score, semantic similarity and theme model matching degree is performed by using the field relevance evaluation layer, and a fifth dimension score is established; A compliance evaluation layer in the six-dimensional evaluation channel is activated, data evaluation based on ethical approval status, regulation compliance check, data privacy protection and academic integrity is performed by using the compliance evaluation layer, and a sixth dimension score is established; The dimension score is established according to the first dimension score, the second dimension score, the third dimension score, the fourth dimension score, the fifth dimension score and the sixth dimension score.

4. The method for screening high quality technological information with self-adapting multi-dimensional evaluation as claimed in claim 1, wherein, The adaptive configuration of dynamic weights of the dimension score mapping includes: The indicators of different dimension scores are normalized to establish a normalized result of the calibration interval; The proportion of the index in all schemes is calculated according to the entropy weight method to establish a first weight result; The subjective weight of the index is obtained to establish a second weight result; After fusion of the first weight result and the second weight result, a dynamic weight is established.

5. A method for screening high quality technological information by self-adapting multi-dimensional evaluation according to claim 4, characterized in that, The calculation of the proportion of the index in all schemes according to the entropy weight method includes: The normalized result is weighted and summed to calculate the proportion value of each index; The specific gravity value is used to calculate an entropy value of each index, and the entropy value represents information uncertainty of the index; An index difference coefficient is calculated according to the entropy value, and the index difference coefficient is used to establish a first weight result.

6. The method for screening high quality technological information with self-adapting multi-dimensional evaluation as claimed in claim 4, wherein, The subjective weight of the index includes: An expert judgment matrix is constructed according to the importance of the index evaluation factors; The expert judgment matrix is subjected to matrix consistency verification; If the consistency verification is passed, the weight of each index is calculated from the expert judgment matrix by using an eigenvector method, and the subjective weight is established.

7. The method for screening high quality technological information with self-adapting multi-dimensional evaluation as claimed in claim 1, wherein, The text feature extraction includes: The preprocessed data set is subjected to TF-IDF vectorization processing, and a term frequency-inverse document frequency matrix is calculated; A semantic vector representation of each word in the preprocessed data set is generated by using a Word2Vec or BERT model; The preprocessed data set is subjected to 1-3-gram syntax feature extraction; A latent topic distribution of the preprocessed data set is extracted based on an LDA model.

8. The method for screening high quality technological information with self-adapting multi-dimensional evaluation as claimed in claim 1, wherein, The network feature calculation includes: Reference network analysis is performed by calculating node centrality indexes; A scholar influence map is constructed, and scholar cooperation relationship and influence analysis is performed; An institution association network is constructed, and cooperation strength between structures is evaluated based on the institution association network.

9. The method for screening high quality technological information with self-adapting multi-dimensional evaluation as claimed in claim 1, wherein, The time sequence feature construction includes: Hot and cold spots in a research field are identified by analyzing publication times of the literature; Time sequence academic influence features of the literature are evaluated according to a citation growth trend of the literature; Time decay analysis is performed by using the time sequence academic influence features, and time effectiveness features of the literature are quantified.

10. A high quality scientific and technical information screening system with adaptive multidimensional evaluation, characterized in that, The system is used for real-time implementation of the adaptive multi-dimensional evaluation high-quality scientific and technological information screening method according to any one of claims 1-9, and the system includes: A multi-source data set establishment module is configured to collect data from a database by using an API interface and a web crawler technology, and establish a multi-source data set; A preprocessed data set establishment module is configured to establish a preprocessed data set by performing data preprocessing on the multi-source data set, and the data preprocessing includes HTML tag cleaning, format unification processing, stop word filtering processing, stem extraction, entity recognition and named entity normalization processing; A multi-dimensional feature set establishment module is configured to establish a standardized multi-dimensional feature set by performing feature extraction on the preprocessed data set, and the feature extraction includes text feature extraction, network feature calculation and time sequence feature construction; A dimension score establishment module is configured to establish a dimension score by performing dimension evaluation of the multi-dimensional feature set under a six-dimensional evaluation channel; A screening result generation module is configured to generate a scientific and technological information screening result by performing weighted calculation after adaptively configuring dynamic weights of the dimension score mapping.

Images