Multi-mode scientific and technological innovation resource data intelligent screening method

Abstract

The invention discloses a multi-mode scientific and technological innovation resource data intelligent screening method. The method comprises the following steps: firstly, setting admission thresholds for research and development investment, the number of researchers, the number of patents and the proportion of new product sales income, so as to preliminarily screen out candidate enterprises meeting the lowest requirements; then, unified preprocessing is conducted on enterprise data from different sources, the multi-source and multi-format problem is solved, and enterprise entity analysis and merging are achieved in combination with methods such as editing distance and cosine similarity; on the basis, text semantic features such as enterprise brief introduction and related news are extracted by means of text mining, word frequency-inverse document frequency and the like, and multi-modal feature vectors are constructed together with numerical features; and finally, performing K-means clustering analysis by adopting comprehensive distance measurement, iteratively calculating a clustering center, and dividing the enterprises into clusters with the highest similarity to obtain a multi-modal clustering result.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for intelligently screening multimodal scientific and technological innovation resource data. Background Art

[0002] As the role of scientific and technological innovation in economic and social development becomes increasingly prominent, how to quickly and accurately identify companies with high innovation potential from massive, multi-source data has become a common need for relevant departments, investment institutions, and industry researchers. Existing technologies generally face the following problems and challenges:

[0003] Data from different databases or institutions often have significant differences in field naming, units, timestamps, currency types, etc., which makes manual cleaning and aggregation labor-intensive and error-prone.

[0004] Usually, the innovation level of an enterprise is measured only by financial indicators (such as revenue, asset size) or simple patent statistics, while ignoring its textual descriptions and research trends in professional fields, making it difficult to fully judge the enterprise's true innovation strength.

[0005] The same company may appear in multiple data sources with different names, spellings or formats. If company entity matching and deduplication are not performed, information redundancy or repeated statistics may easily occur.

[0006] Traditional methods mainly rely on manual retrieval, manual screening and partially automated sorting, and are unable to fully utilize clustering, text mining and other technologies to conduct in-depth analysis of large-scale, multimodal data. Summary of the Invention

[0007] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a method for intelligent screening of multimodal scientific and technological innovation resource data, comprising:

[0008] Admission indicator setting: Based on enterprise scale and innovation requirements, define the minimum admission threshold function for enterprise R&D investment, number of patents, number of scientific researchers, and proportion of new product sales revenue;

[0009] Data unification: Enterprise records from multiple data sources with inconsistent formats are standardized through unification functions, and different records of the same enterprise are deduplicated and merged using entity matching algorithms.

[0010] Text feature extraction: Perform natural language processing on text-containing data such as corporate bulletins, official websites, and technical documents, calculate the word frequency-inverse document frequency vector, and obtain the corporate text feature representation;

[0011] Multimodal clustering analysis: After fusing the numerical and textual features of an enterprise, a K-means clustering method based on a comprehensive distance metric is used. Numerical features are measured using Euclidean distance, and textual features are measured using semantic similarity. Adjustable weights are set to balance numerical and textual information.

[0012] Enterprise screening and output: After the clustering and grouping, the enterprises are judged based on the access indicators to see whether they meet the threshold requirements, and the target enterprises that meet all the minimum access standards are screened out and the result list is output.

[0013] In a preferred embodiment of the present invention, the admission index setting includes:

[0014] Several key quantitative indicators are defined for each company x to assess its scale and degree of innovation;

[0015] Company size:

[0016] R&D investment x and the number of researchers N x , used to reflect the scale of the company in the field of scientific and technological innovation;

[0017] Innovation output:

[0018] Number of patents P x and new product sales revenue share x , using this as an indicator to reflect the innovation output of enterprises;

[0019] Innovation areas:

[0020] Define a field set Fields = {F1, F2, ...., F m}, each F q Represents a technical field, q=1,2,....,m;

[0021] If company x is in field F q Inside, then F q =1, otherwise 0;

[0022] m is the total number of fields;

[0023] A threshold function is defined for each indicator to determine whether the company meets the minimum entry standard. v When compared, the results are as follows:

[0024]

[0025] Among them, T v (v) is the threshold function of indicator v, θ v is the threshold value of the indicator v;

[0026] If Rx ≥R min Then T R (R x )=1, meets the admission criteria, otherwise 0;

[0027] If N x ≥N min Then T N (N x )=1, meets the admission criteria, otherwise 0;

[0028] By analogy, based on P x ≥P min and I x ≥I min The number of patents P x and new product sales revenue share x Set T P (P x ) and T I (I x );

[0029] R min is the threshold of R&D investment;

[0030] N min is the threshold number of scientific researchers;

[0031] P min is the threshold number of patents;

[0032] I min is the threshold of the proportion of new product sales revenue;

[0033] These thresholds are set based on policy needs or data quantiles;

[0034] Define the entry function E(x) of company x:

[0035] E(x)=T R (R x )∧T N (N x )∧T P (P x )∧T I (I x );

[0036] Among them, ∧ is the AND operation;

[0037] T R (R x ) is the threshold function of R&D investment;

[0038] T N (N x ) is the threshold function of the number of scientific researchers;

[0039] T P (P x ) is the threshold function of the number of patents;

[0040] T I (I x ) is the threshold function of the proportion of new product sales revenue;

[0041] Only when all indicators are greater than or equal to the threshold, that is, E(x) = 1, can the company be selected.

[0042] In a preferred embodiment of the present invention, the data unification processing includes:

[0043] In multi-source data integration scenarios, different data sources may have problems such as different field names, inconsistent units, and confusing formats;

[0044] Therefore, a unification function U is introduced to map the records of each source into a standardized pattern. If x src From a source, then:

[0045] U(x src ) = x′;

[0046] Among them, U(*) is a unification function that maps company data records from various sources and formats to a unified standard model;

[0047] x src Represents a set of company records obtained from a data source;

[0048] x′ is the unified representation of the company data after normalization. The unified model includes company name, location, industry, revenue, number of employees, number of patents, and number of papers. If some sources provide revenue in millions of US dollars and others in local currency, they are uniformly converted to the same currency unit in U.

[0049] In a preferred embodiment of the present invention, the text feature extraction includes:

[0050] Company-related attributes contain a large amount of text, and informative features are extracted through NLP methods:

[0051] Calculate term frequency-inverse document frequency for a text document:

[0052] tfidf(τ,dc,D)=tf(τ,dc)·idf(τ,D);

[0053] Among them, tfidf(*) is the term frequency-inverse document frequency, which is a numerical indicator used to measure the importance of a word in a specific document;

[0054] τ is a specific keyword or term;

[0055] dc is a specific single document;

[0056] D is the corpus consisting of all documents;

[0057] tf(τ,dc) is the frequency of word τ in document d;

[0058] idf(τ,D) is the rarity of τ in the corpus. The fewer documents that contain the word, the larger the value.

[0059] in, It emphasizes words that are high frequency in a specific document but rare in the entire corpus, which can highlight the company's focus areas and ultimately map each document into a feature vector in the word frequency space;

[0060] Where |D| is the total number of documents in the corpus;

[0061] |{dc′∈D:τ∈dc′}| represents the number of documents containing the keyword τ, which is used to measure the prevalence of the keyword in the entire document set;

[0062] m1 and m2 are non-negative integers used to reduce the impact of extreme values and perform smoothing. m1 is used to make small increments on the numerator, and m2 is used to prevent the denominator from being 0 or a very small value, thus avoiding numerical instability.

[0063] dc′ is the low-frequency document in the entire corpus.

[0064] In a preferred embodiment of the present invention, the multimodal cluster analysis includes:

[0065] When collecting data, the information of the same company often appears in multiple data sources. After identifying the same company, it is necessary to map the data from each source to a unified model and define the mapping function M. s (*) The conflicting data x from source s s ' is converted to a standard field:

[0066] x″=M s (x s ″);

[0067] Where x″ is the standardized company data after resolving conflicts among various data sources;

[0068] M s (*) is a mapping function used to select the data with the highest credibility as the final retained value when the data conflict;

[0069] x s ″ is the data provided by source s, and there is data from other sources that conflicts with it;

[0070] Finally, we get the company entity list {e1,e2,....,e X}, each e i Aggregates the company's information from various sources, where i = 1, 2, ...., X, X is the total number of corporate entities, and for each corporate entity e i Generate unified company profiles;

[0071] For a data set χ consisting of X company entities, we have:

[0072] χ={e1,e2,....,e X};

[0073] For each corporate entity i Separate the features into two parts:

[0074] Numerical features Including revenue, number of employees, and number of patents;

[0075] Text embedding features The model extracts company profiles and related news;

[0076] In order to group companies according to similarity, a clustering function Cl(e) can be defined to divide all entities into one of k clusters {C1, C2, ..., C k};

[0077] Combined with K-means clustering:

[0078] Divide the companies into k clusters, with the goal of minimizing the intra-cluster variance;

[0079] Iteratively solve the central vector μ r , used to discover similar groups;

[0080] Randomly select z company entities as initial centers in:

[0081]

[0082] r=1,2,....,z,s r is the initial selected index;

[0083] is the initially selected numerical feature, is the initially selected text feature;

[0084] Introduce the fusion cross-modal correlation and for each company entity e i Calculate all center vectors Distance:

[0085]

[0086] in, After the tth iteration, the company entity e i To the rth cluster center The comprehensive distance;

[0087] is the center vector of the rth cluster after t iterations;

[0088] is the center vector of the rth cluster in the numerical space after the tth iteration;

[0089] is the center vector of the rth cluster in the text space after the tth iteration;

[0090] ||*||2 represents the Euclidean norm;

[0091] is the relevance weight, which controls the influence of cross-modal similarity on the overall distance;

[0092] It is the cross-modal similarity. The closer the value is to 1, the more consistent the entity is with the central cluster in terms of numerical and textual relationship. It is calculated as follows:

[0093]

[0094] Among them, Δ cm is the cross-modal difference, Δ cm =φ1Δ cos +φ2Δ κ +φ3Δ dist ,Through negative exponential mapping, multiple sub-difference measures are introduced to measure the ,internal correlation between entities and cluster centers in these two modalities, which is used to ,solve the problem of mismatch between text and numerical data when filtering ,data;

[0095] φ is the attenuation coefficient, the larger the value, the more sensitive it is to the difference;

[0096] φ1, φ2, and φ3 are adjustable weights used to weigh the importance of different sub-difference measures;

[0097] Among them, Δ cos is the cosine similarity difference,

[0098] cos(*) represents the cosine similarity of two vectors;

[0099] |*| means taking the absolute value;

[0100] Δ κ is the Pearson correlation coefficient difference,

[0101] κ(*) is the Pearson correlation coefficient, which is used to measure the linear correlation between value and text;

[0102] Δ dist is the dimension distribution difference,

[0103] express The mean value of the numeric vector in dimension o;

[0104] express The mean value of the text vector in dimension o;

[0105] express The mean value of the numeric vector in dimension o;

[0106] express The mean value of the text vector in dimension o;

[0107] is a dynamic weight function, which is calculated as follows:

[0108]

[0109] Among them, σ(*) is the Logistic activation function, which is used to map real numbers to the value range of (0,1);

[0110] γ is the difference weight, which is used to determine the sensitivity of the difference between numerical distance and text distance;

[0111] λ is the dispersion weight, which is used to weigh the difference in the dispersion of the cluster in the two modes;

[0112] η low quality correction, additional correction of weight when the text reliability is low;

[0113] is the intra-cluster variance, is the intra-cluster text variance, when When , it means that the cluster is more compact in the text dimension and tends to focus on text distance. On the contrary, it means that the cluster is more compact in the numerical dimension and tends to focus on numerical distance.

[0114] υ i is the text reliability, (1-υ i ) is the text reliability correction item. When the text quality is low, the system will reduce the weight of the text distance to reduce the impact of noise on clustering;

[0115] e i Assign to the nearest cluster:

[0116]

[0117] in, is the rth cluster after the tth iteration;

[0118] e i :r means to put all the i For instance, the entity with the smallest distance to the cluster index is r, and all of them are included in the cluster among;

[0119] Indicates that among all cluster centers, the center vector that minimizes the distance function is selected The corresponding index r;

[0120] For each cluster Take the mean in the numerical and text spaces to update the centroid for the calculation of the comprehensive distance in the next iteration:

[0121]

[0122]

[0123] in, is the numerical center of the rth cluster after the t+1th iteration;

[0124] is the text center of the rth cluster after the t+1th iteration;

[0125] It is a cluster the number of corporate entities in;

[0126] It means taking the numerical vectors of all entities in the cluster and adding them one by one, and finally dividing them by the number of cluster members to get the numerical center;

[0127] It means taking the text vectors of all entities in the cluster and adding them one by one, and finally dividing them by the number of cluster members to get the text center;

[0128] Indicates that it is assigned to a cluster Corporate entities in i ;

[0129] The cluster centers are updated in their respective dimensions, which is equivalent to taking the average in the numerical space and the text space respectively;

[0130] Stop if the differences between all new and old centers are small enough, or if the maximum number of iterations has been reached:

[0131]

[0132] Among them, max 1≤r≤k (*) represents the maximum change of the center vectors of all clusters between the old and new iterations; when all changes are no greater than the convergence threshold ε, it means that the center positions of all clusters have stabilized;

[0133] is the center vector of the rth cluster after t+1 iterations;

[0134] ||*||2 is the Euclidean norm;

[0135] ε is the pre-set convergence threshold, otherwise return to the allocation step;

[0136] Output the final cluster partition {C1,C2,.....,C k} and multimodal centroids {μ1,μ2,.....,μ k};

[0137] The clustering results help organize and display companies by category or similarity on the platform.

[0138] The present invention also discloses a computer system, comprising:

[0139] processor;

[0140] a memory for storing processor-executable instructions;

[0141] Among them, the processor is configured to implement the method for intelligent screening of multimodal scientific and technological innovation resource data when executing the executable instructions.

[0142] The present invention also discloses a computer-readable storage medium, comprising:

[0143] a memory having a computer program stored thereon;

[0144] A processor is used to execute the program in the memory to implement the method for intelligent screening of multimodal scientific and technological innovation resource data as described in the rights.

[0145] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0146] Unified multi-source data integration: By introducing unified functions and company entity matching technology, enterprise data from different sources and formats are mapped to a unified standard model, achieving effective merging and deduplication of multi-source data, and ensuring data integrity and consistency.

[0147] Multimodal feature mining: Not only does it use structured indicators such as enterprise size, number of patents, and number of research papers, but it also uses NLP (natural language processing) methods to extract key information from texts such as company profiles and news reports, enriching the dimensions of enterprise innovation output evaluation.

[0148] Threshold screening and flexible parameter adjustment: By setting entry thresholds for key indicators such as revenue, number of employees, number of patents, and proportion of new product sales revenue, and flexibly adjusting them based on policy requirements or data quantiles, we ensure that the screened companies meet the expected standards in terms of scale and innovation.

[0149] Intelligent clustering and visual analysis: Combining multimodal distance metrics (numeric + text) and K-means clustering, companies are classified according to similarity or specific innovation characteristics, helping managers or researchers better understand market structure and quickly identify the most promising innovative companies.

[0150] Through the above technical means, the present invention not only improves the accuracy and efficiency of enterprise screening, but also provides strong technical support for the discovery, evaluation and industry research of innovative enterprises.

[0151] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0152] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0153] Figure 1 It is a flow chart of the method of the present invention.

[0154] Figure 2 It is a flowchart of iterative calculation of multimodal clustering. DETAILED DESCRIPTION

[0155] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0156] like Figure 1 As shown, the present invention discloses a method for intelligent screening of multimodal scientific and technological innovation resource data, comprising:

[0157] 1. Data screening objectives and screening criteria

[0158] Several key quantitative indicators are defined for each company x to assess its scale and degree of innovation;

[0159] Company size:

[0160] R&D investment x and the number of researchers N x , used to reflect the scale of the company in the field of scientific and technological innovation;

[0161] Innovation output:

[0162] Number of patents P x and new product sales revenue share x , using this as an indicator to reflect the innovation output of enterprises;

[0163] Innovation areas:

[0164] Define a field set Fields = {F1, F2, ...., F m}, each F q Represents a technical field, q=1,2,....,m;

[0165] If company x is in field F q Inside, then F q =1, otherwise 0;

[0166] m is the total number of fields.

[0167] 1.1 Entry Threshold

[0168] A threshold function is defined for each indicator to determine whether the company meets the minimum entry standard. v When compared, the results are as follows:

[0169]

[0170] Among them, T v (v) is the threshold function of indicator v, θ v is the threshold value of the indicator v;

[0171] If R x ≥R min Then T R (R x )=1, meets the admission criteria, otherwise 0;

[0172] If N x ≥N min Then T N (N x )=1, meets the admission criteria, otherwise 0;

[0173] By analogy, based on P x ≥P min and I x ≥Imin The number of patents P x and new product sales revenue share x Set T P (P x ) and T I (I x );

[0174] R min is the threshold of R&D investment;

[0175] N min is the threshold number of scientific researchers;

[0176] P min is the threshold number of patents;

[0177] I min is the threshold of the proportion of new product sales revenue;

[0178] These thresholds are set based on policy requirements or data quantiles.

[0179] 1.2 Screening Rules

[0180] Define the entry function E(x) of company x:

[0181] E(x)=T R (R x )∧T N (N x )∧T P (P x )∧T I (I x );

[0182] Among them, ^ is the AND operation;

[0183] T R (R x ) is the threshold function of R&D investment;

[0184] T N (N x ) is the threshold function of the number of scientific researchers;

[0185] T P (P x ) is the threshold function of the number of patents;

[0186] T I (I x ) is the threshold function of the proportion of new product sales revenue;

[0187] Only when all indicators are greater than or equal to the threshold, that is, E(x) = 1, can the company be selected.

[0188] 2. Data Preprocessing

[0189] 2.1 Format Unification

[0190] In a multi-source data integration scenario, different data sources may have problems such as different field naming, inconsistent units, and confusing formats.

[0191] Therefore, a unification function U is introduced to map the records of each source into a standardized pattern. If x src From a source, then:

[0192] U(x src ) = x′;

[0193] Among them, U(*) is a unification function, which converts or maps company data records from various sources and formats into a unified standard model.

[0194] x src Represents a set of company records obtained from a data source;

[0195] x′ is the unified representation of the company data after normalization. The unified model includes fields such as company name, location, industry, revenue, number of employees, number of patents, and number of papers. If some sources provide revenue in millions of US dollars and others in local currency, they are uniformly converted to the same currency unit in U (for example, uniformly converted to US dollars).

[0196] 2.2 Text Feature Extraction

[0197] Company-related attributes contain a large amount of text (company profiles, news reports, etc.). Informative features are extracted through NLP methods:

[0198] Calculate the term frequency-inverse document frequency (TF–IDF) vector for text documents (such as company introductions and press releases):

[0199] tfidf(τ,dc,D)=tf(τ,dc)·idf(τ,D);

[0200] Among them, tfidf(*) is the term frequency-inverse document frequency, which is a numerical indicator used to measure the importance of a word in a specific document;

[0201] τ is a specific keyword or term;

[0202] dc is a specific single document;

[0203] D is the corpus consisting of all documents;

[0204] tf(τ,dc) is the frequency of word τ in document d;

[0205] idf(τ,D) is the rarity of τ in the corpus. The fewer documents that contain the word, the larger the value.

[0206] in, It emphasizes words that are high frequency in a specific document but rare in the entire corpus, which can highlight the company's focus areas and ultimately map each document into a feature vector in the word frequency space;

[0207] Where |D| is the total number of documents in the corpus;

[0208] |{dc′∈D:τ∈dc′}| represents the number of documents containing the keyword τ, which is used to measure the prevalence of the keyword in the entire document set;

[0209] m1 and m2 are non-negative integers used to reduce the impact of extreme values and perform smoothing. m1 is used to make small increments on the numerator, and m2 is used to prevent the denominator from being 0 or a very small value, thus avoiding numerical instability.

[0210] dc′ is the low-frequency document in the entire corpus.

[0211] By calculating word frequency and inverse document frequency, words that appear frequently in a single document but are relatively rare in the entire corpus can be highlighted, helping the system discover key information that best represents a company's core technology areas or business priorities.

[0212] It can quickly filter out irrelevant or highly common words (such as "the", "company", etc.), highlight the keywords that can truly distinguish the technical directions of different companies, and facilitate the system to efficiently search, compare and cluster texts from multiple companies.

[0213] 3. Data fusion and standardization

[0214] 3.1 Cross-source company entity matching

[0215] Information about the same company may appear in multiple data sources with slightly different names or formats. To merge them into the same company entity, entity resolution (record linking) is required:

[0216] If both descriptions contain highly similar keywords such as "AI-driven healthcare," the cosine similarity will be close to 1. Through a comprehensive assessment of string and vector similarity, it is determined that they point to the same real company.

[0217] 3.1 Mode Mapping:

[0218] When collecting data, the information of the same company often appears in multiple data sources. After identifying the same company, it is necessary to map the data from each source to a unified model and define the mapping function M. s (*) The conflicting data x from source ss ' is converted to a standard field:

[0219] x″=M s (x s ″);

[0220] Where x″ is the standardized company data after resolving conflicts among various data sources;

[0221] M s (*) is a mapping function used to select the data with the highest credibility as the final retained value when the data conflict;

[0222] x s ″ is the data provided by source s, and there is data from other sources that conflicts with it;

[0223] Finally, we get the company entity list {e1,e2,....,e X}, each e i Aggregates the company's information from various sources, where i = 1, 2, ...., X, X is the total number of corporate entities, and for each corporate entity e i Generate a unified company profile.

[0224] 4. Intelligent screening model

[0225] like Figure 2 As shown in the figure, after obtaining the standardized features, the clustering model is used to screen and present the companies;

[0226] 4.1 Clustering and Segmentation

[0227] For a data set χ consisting of X company entities, we have:

[0228] χ={e1,e2,....,e X};

[0229] For each corporate entity i Separate the features into two parts:

[0230] Numerical features Including revenue, number of employees, and number of patents;

[0231] Text embedding features The model extracts company profiles and related news;

[0232] 4.2 Multimodal distance metrics

[0233] Define the comprehensive distance function d(*) of "value + text";

[0234] In order to group companies according to similarity, a clustering function Cl(e) can be defined to divide all entities into one of k clusters {C1, C2, ..., Ck};

[0235] Combined with K-means clustering:

[0236] Divide the companies into k clusters, with the goal of minimizing the intra-cluster variance;

[0237] Iteratively solve the central vector μ r , used to discover similar groups;

[0238] Used to discover similar groups, such as "cluster of large enterprises" or "cluster of small startups with high patent counts";

[0239] 4.3 Iteration Process

[0240] 4.3.1 Initialization

[0241] Randomly select z company entities as initial centers in:

[0242]

[0243] r=1,2,....,z,s r is the initial selected index;

[0244] is the initially selected numerical feature, is the initially selected text feature.

[0245] 4.3.2 Allocation

[0246] Introduce the fusion cross-modal correlation and for each company entity e i Calculate all center vectors Distance:

[0247]

[0248] in, After the tth iteration, the company entity e i To the rth cluster center The comprehensive distance;

[0249] is the center vector of the rth cluster after t iterations;

[0250] is the center vector of the rth cluster in the numerical space after the tth iteration;

[0251] is the center vector of the rth cluster in the text space after the tth iteration;

[0252] ||*||2 represents the Euclidean norm;

[0253] is the relevance weight, which controls the influence of cross-modal similarity on the overall distance;

[0254] It is the cross-modal similarity. The closer the value is to 1, the more consistent the entity is with the central cluster in terms of numerical and textual relationship. It is calculated as follows:

[0255]

[0256] Among them, Δ cm is the cross-modal difference, Δ cm =φ1Δ cos +φ2Δ κ +φ3Δ dist ,Through negative exponential mapping, multiple sub-difference measures are introduced to measure the ,internal correlation between entities and cluster centers in these two modalities, which is used to ,solve the problem of mismatch between text and numerical data when filtering ,data;

[0257] φ is the attenuation coefficient, the larger the value, the more sensitive it is to the difference;

[0258] φ1, φ2, and φ3 are adjustable weights used to weigh the importance of different sub-difference measures;

[0259] Among them, Δ cos is the cosine similarity difference,

[0260] cos(*) represents the cosine similarity of two vectors;

[0261] |*| means taking the absolute value;

[0262] Δ κ is the Pearson correlation coefficient difference,

[0263] κ(*) is the Pearson correlation coefficient, which is used to measure the linear correlation between value and text;

[0264] Δ dist is the dimension distribution difference,

[0265] express The mean value of the numeric vector in dimension o;

[0266] express The mean value of the text vector in dimension o;

[0267] express The mean value of the numeric vector in dimension o;

[0268] express The mean value of the text vector in dimension o;

[0269] is a dynamic weight function, which is calculated as follows:

[0270]

[0271] Among them, σ(*) is the Logistic activation function, which is used to map real numbers to the value range of (0,1);

[0272] γ is the difference weight, which is used to determine the sensitivity of the difference between numerical distance and text distance;

[0273] λ is the dispersion weight, which is used to weigh the difference in the dispersion of the cluster in the two modes;

[0274] η low quality correction, additional correction of weight when the text reliability is low;

[0275] is the intra-cluster variance, is the intra-cluster text variance, when When , it means that the cluster is more compact in the text dimension and tends to focus on text distance. On the contrary, it means that the cluster is more compact in the numerical dimension and tends to focus on numerical distance.

[0276] υ i is the text reliability, (1-υ i ) is the text reliability correction item. When the text quality is low, the system will reduce the weight of the text distance to reduce the impact of noise on clustering;

[0277] e i Assign to the nearest cluster:

[0278]

[0279] in, is the rth cluster after the tth iteration;

[0280] e i :r means to put all the i For instance, the entity with the smallest distance to the cluster index is r, and all of them are included in the cluster among;

[0281] Indicates that among all cluster centers, the center vector that minimizes the distance function is selected The corresponding index r.

[0282] 4.3.3 Update

[0283] For each cluster Take the mean in the numerical and text spaces to update the centroid for the calculation of the comprehensive distance in the next iteration:

[0284]

[0285] in, is the numerical center of the rth cluster after the t+1th iteration;

[0286] is the text center of the rth cluster after the t+1th iteration;

[0287] It is a cluster the number of corporate entities in;

[0288] It means taking the numerical vectors of all entities in the cluster and adding them one by one, and finally dividing them by the number of cluster members to get the numerical center;

[0289] It means taking the text vectors of all entities in the cluster and adding them one by one, and finally dividing them by the number of cluster members to get the text center;

[0290] Indicates that it is assigned to a cluster Corporate entities in i ;

[0291] The cluster centers are updated with mean values in their respective dimensions, which is equivalent to taking the average in the numerical space and the text space respectively.

[0292] 4.3.4 Convergence determination

[0293] Stop if the differences between all new and old centers are small enough, or if the maximum number of iterations has been reached:

[0294]

[0295] Among them, max 1≤r≤k (*) represents the maximum change of the center vectors of all clusters between the old and new iterations; when all changes are no greater than the convergence threshold ε, it means that the center positions of all clusters have stabilized;

[0296] is the center vector of the rth cluster after t+1 iterations;

[0297] ||*||2 is the Euclidean norm;

[0298] ε is a pre-set convergence threshold, otherwise return to the allocation step.

[0299] 4.3.5 Output

[0300] Output the final cluster partition {C1,C2,.....,C k} and multimodal centroids {μ1,μ2,.....,μ k};

[0301] The clustering results help organize and display companies by category or similarity on the platform.

[0302] 5. Continuous Learning and Feedback

[0303] After the system is launched, user feedback, such as platform visits and professional reviews of company listings, can be continuously collected to supplement the training set with additional labels. Online evaluation can also be conducted to measure user clicks or usage of the recommended company list, and further optimization can be achieved using multi-armed bandit or reinforcement learning methods. However, offline metrics such as precision and recall remain the core, ensuring high recall (not missing truly promising companies) and high precision (reducing noise companies).

[0304] By continuously standardizing, integrating, and updating public and trusted data sources (such as Crunchbase, patent databases, and paper databases), and applying advanced AI models to screen and rank companies, our system reliably identifies companies that meet innovation criteria, providing a high-value platform for the public to explore technology startups. The entire design is modular and scalable, allowing for the integration of new data sources and algorithm upgrades at any time. Combined with detailed evaluation and tuning, this ensures the system's superior effectiveness and broad applicability in screening innovative companies.

[0305] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A method for intelligent screening of multimodal scientific and technological innovation resource data, characterized in that: include: Admission indicator setting: Based on enterprise scale and innovation requirements, define the minimum admission threshold function for enterprise R&D investment, number of patents, number of scientific researchers, and proportion of new product sales revenue; Data unification: Enterprise records from multiple data sources with inconsistent formats are standardized through unification functions, and different records of the same enterprise are deduplicated and merged using entity matching algorithms. Text feature extraction: Perform natural language processing on text-containing data such as corporate bulletins, official websites, and technical documents, calculate the word frequency-inverse document frequency vector, and obtain the corporate text feature representation; Multimodal clustering analysis: After fusing the numerical and textual features of the enterprise, a K-means clustering method based on comprehensive distance measurement is used; Numerical features are measured by Euclidean distance, textual features are measured by semantic similarity, and adjustable weights are set to balance numerical and textual information. Enterprise screening and output: After the clustering and grouping, the enterprises are judged based on the access indicators to see whether they meet the threshold requirements, and the target enterprises that meet all the minimum access standards are screened out and the result list is output.

2. The method for intelligent screening of multimodal scientific and technological innovation resource data according to claim 1, characterized in that: The admission indicator setting includes: Several key quantitative indicators are defined for each company x to assess its scale and degree of innovation; Company size: R&D investment x and the number of researchers N x , used to reflect the scale of the company in the field of scientific and technological innovation; Innovation output: Number of patents P x and new product sales revenue share x , using this as an indicator to reflect the innovation output of enterprises; Innovation areas: Define a field set Fields = {F1, F2, ...., F m }, each F q Represents a technical field, q=1,2,....,m; If company x is in field F q Inside, then F q =1, otherwise 0; m is the total number of fields; A threshold function is defined for each indicator to determine whether the company meets the minimum entry standard. v When compared, the results are as follows: Among them, T v (v) is the threshold function of indicator v, θ v is the threshold value of the indicator v; If R x ≥R min Then T R (R x )=1, meets the admission criteria, otherwise 0; If N x ≥N min Then T N (N x )=1, meets the admission criteria, otherwise 0; By analogy, based on P x ≥P min and I x ≥I min The number of patents P x and new product sales revenue share x Set T P (P x ) and T I (I x ); R min is the threshold of R&D investment; N min is the threshold number of scientific researchers; P min is the threshold number of patents; I min is the threshold of the proportion of new product sales revenue; These thresholds are set based on policy needs or data quantiles; Define the entry function E(x) of company x: E(x)=T R (R x )∧T N (N x )∧T P (P x )∧T I (I x ); Among them, ∧ is the AND operation; T R (R x ) is the threshold function of R&D investment; T N (N x ) is the threshold function of the number of scientific researchers; T P (P x ) is the threshold function of the number of patents; T I (I x ) is the threshold function of the proportion of new product sales revenue; Only when all indicators are greater than or equal to the threshold, that is, E(x)=1, can the company be selected.

3. The method for intelligent screening of multimodal scientific and technological innovation resource data according to claim 1, characterized in that: The data unified processing includes: In multi-source data integration scenarios, different data sources may have problems such as different field names, inconsistent units, and confusing formats; Therefore, a unification function U is introduced to map the records of each source into a standardized pattern. If x src From a source, then: U(x src )=x′; U(*) is a unification function that maps company data records from various sources and formats to a unified standard model. x src Represents a set of company records obtained from a data source; x′ is the unified representation of the company data after normalization. The unified model includes company name, location, industry, revenue, number of employees, number of patents, and number of papers. If some sources provide revenue in millions of US dollars and others in local currency, they are uniformly converted to the same currency unit in U.

4. The method for intelligent screening of multimodal scientific and technological innovation resource data according to claim 1, characterized in that: The text feature extraction includes: Company-related attributes contain a large amount of text, and informative features are extracted through NLP methods: Calculate term frequency-inverse document frequency for a text document: tfidf(τ,dc,D)=tf(τ,dc)·idf(τ,D); Among them, tfidf(*) is the term frequency-inverse document frequency, which is a numerical indicator used to measure the importance of a word in a specific document; τ is a specific keyword or term; dc is a specific single document; D is the corpus consisting of all documents; tf(τ,dc) is the frequency of word τ in document d; idf(τ,D) is the rarity of τ in the corpus. The fewer documents that contain the word, the larger the value. in, It emphasizes words that are high frequency in a specific document but rare in the entire corpus, which can highlight the company's focus areas and ultimately map each document into a feature vector in the word frequency space; Where |D| is the total number of documents in the corpus; |{dc′∈D:τ∈dc′}| represents the number of documents containing the keyword τ, which is used to measure the prevalence of the keyword in the entire document set; m1 and m2 are non-negative integers used to reduce the impact of extreme values and perform smoothing. m1 is used to make small increments on the numerator, and m2 is used to prevent the denominator from being 0 or a very small value, thus avoiding numerical instability. dc′ is the low-frequency document in the entire corpus.

5. The method for intelligent screening of multimodal scientific and technological innovation resource data according to claim 4, characterized in that: The multimodal cluster analysis includes: When collecting data, the information of the same company often appears in multiple data sources. After identifying the same company, it is necessary to map the data from each source to a unified model and define the mapping function M. s (*) The conflicting data x from source s s ' is converted to a standard field: x″=M s (x s ″); Where x″ is the standardized company data after resolving conflicts among various data sources; M s (*) is a mapping function used to select the data with the highest credibility as the final retained value when the data conflict; x s ″ is the data provided by source s, and there is data from other sources that conflicts with it; Finally, we get the company entity list {e1,e2,....,e X }, each e i Aggregates the company's information from various sources, where i = 1, 2, ...., X, X is the total number of corporate entities, and for each corporate entity e i Generate unified company profiles; For a data set χ consisting of X company entities, we have: χ={e1,e2,....,e X }; For each corporate entity i Separate the features into two parts: Numerical features Including revenue, number of employees, and number of patents; Text embedding features The model extracts company profiles and related news; In order to group companies according to similarity, a clustering function Cl(e) can be defined to divide all entities into one of k clusters {C1, C2, ..., C k }; Combined with K-means clustering: Divide the companies into k clusters, with the goal of minimizing the intra-cluster variance; Iteratively solve the central vector μ r , used to discover similar groups; Randomly select z company entities as initial centers in: r=1,2,....,z,s r is the initial selected index; is the initially selected numerical feature, is the initially selected text feature; Introduce the fusion cross-modal correlation and for each company entity e i Calculate all center vectors Distance: in, After the tth iteration, the company entity e i To the rth cluster center The comprehensive distance; is the center vector of the rth cluster after t iterations; is the center vector of the rth cluster in the numerical space after the tth iteration; is the center vector of the rth cluster in the text space after the tth iteration; ||*||2 represents the Euclidean norm; is the relevance weight, which controls the influence of cross-modal similarity on the overall distance; It is the cross-modal similarity. The closer the value is to 1, the more consistent the entity is with the central cluster in terms of numerical and textual relationship. It is calculated as follows: Among them, Δ cm is the cross-modal difference, Δ cm =φ1Δ cos +φ2Δ κ +φ3Δ dist Through negative exponential mapping, multiple sub-difference measures are introduced to measure the "internal correlation" between entities and cluster centers in these two modes, which is used to solve the problem of mismatch between text and numerical data when filtering data; φ is the attenuation coefficient, the larger the value, the more sensitive it is to the difference; φ1, φ2, and φ3 are adjustable weights used to weigh the importance of different sub-difference measures; Among them, Δ cos is the cosine similarity difference, cos(*) represents the cosine similarity of two vectors; |*| means taking the absolute value; Δ κ is the Pearson correlation coefficient difference, κ(*) is the Pearson correlation coefficient, which is used to measure the linear correlation between value and text; Δ dist is the dimension distribution difference, express The mean value of the numeric vector in dimension o; express The mean value of the text vector in dimension o; express The mean value of the numeric vector in dimension o; express The mean value of the text vector in dimension o; is a dynamic weight function, which is calculated as follows: Among them, σ(*) is the Logistic activation function, which is used to map real numbers to the value range of (0,1); γ is the difference weight, which is used to determine the sensitivity of the difference between numerical distance and text distance; λ is the dispersion weight, which is used to weigh the difference in the dispersion of the cluster in the two modes; η low quality correction, additional correction of weight when the text reliability is low; is the intra-cluster variance, is the intra-cluster text variance, when When , it means that the cluster is more compact in the text dimension and tends to focus on text distance. On the contrary, it means that the cluster is more compact in the numerical dimension and tends to focus on numerical distance. υ i is the text reliability, (1-υ i ) is the text reliability correction item. When the text quality is low, the system will reduce the weight of the text distance to reduce the impact of noise on clustering; e i Assign to the nearest cluster: in, is the rth cluster after the tth iteration; e i :r means to put all the i For instance, the entity with the smallest distance to the cluster index is r, and all of them are included in the cluster among; Indicates that among all cluster centers, the center vector that minimizes the distance function is selected The corresponding index r; For each cluster Take the mean in the numerical and text spaces to update the centroid for the calculation of the comprehensive distance in the next iteration: in, is the numerical center of the rth cluster after the t+1th iteration; is the text center of the rth cluster after the t+1th iteration; It is a cluster the number of corporate entities in; It means taking the numerical vectors of all entities in the cluster and adding them one by one, and finally dividing them by the number of cluster members to get the numerical center; It means taking the text vectors of all entities in the cluster and adding them one by one, and finally dividing them by the number of cluster members to get the text center; Indicates that it is assigned to a cluster Corporate entities in i ; The cluster centers are updated in their respective dimensions, which is equivalent to taking the average in the numerical space and the text space respectively; Stop if the differences between all new and old centers are small enough, or if the maximum number of iterations has been reached: Among them, max 1≤r≤k (*) represents the maximum change of the center vectors of all clusters between the old and new iterations; when all changes are no greater than the convergence threshold ε, it means that the center positions of all clusters have stabilized; is the center vector of the rth cluster after t+1 iterations; ||*||2 is the Euclidean norm; ε is the pre-set convergence threshold, otherwise return to the allocation step; Output the final cluster partition {C1,C2,.....,C k } and multimodal centroids {μ1,μ2,.....,μ k }; The clustering results help organize and display companies by category or similarity on the platform.

6. A computer system, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement a multimodal scientific and technological innovation resource data intelligent screening method according to any one of claims 1 to 5 when executing the executable instructions.

7. A computer-readable storage medium, characterized in that include: a memory having a computer program stored thereon; A processor is used to execute the program in the memory to implement a multimodal scientific and technological innovation resource data intelligent screening method according to any one of claims 1 to 5.

Images