Internet big data analysis method and system based on artificial intelligence

Abstract

The invention provides an Internet big data analysis method and system based on artificial intelligence, and the method comprises the steps: collecting heterogeneous data of different data sources from the medical equipment industry, carrying out the standardization processing, carrying out the key feature extraction of the data after the standardization processing, and carrying out the fusion into a unified comprehensive feature vector; generating a dynamically updated user portrait according to the comprehensive feature vector; a dynamically updated knowledge graph is constructed by combining the medical instrument industry with the dynamically updated user portrait, so that the market change is kept in the latest state; market analysis is carried out by using the knowledge graph, and product demand prediction and user recommendation are provided; and providing a personalized marketing scheme according to the comprehensive market strategy suggestions, and performing real-time optimization. Through fusion processing of the multi-dimensional data and construction of the industry knowledge graph, comprehensive capturing and dynamic updating of user requirements are achieved, intelligent market decision support based on the knowledge graph is provided, and the limitation of the prior art is overcome.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular to an Internet big data analysis method and system based on artificial intelligence. Background Art

[0002] With the rapid development of the medical device industry, the production, sales and circulation of medical devices have become increasingly complex. The medical device industry covers multiple entities such as hospitals, clinics, pharmacies, and suppliers. The needs of these entities vary. With the intensification of market competition and policy changes, how to effectively conduct market insights, user management, and precision marketing have become key challenges for major medical device companies. In this context, intelligent systems based on the Internet and big data technologies have gradually become an effective means to optimize processes and improve efficiency. However, the existing Internet big data analysis system still has many shortcomings when facing the unique needs of the medical device industry.

[0003] First, a major problem with existing technologies is the limitation of data sources. Most big data analysis systems can only process single-dimensional or structured data, such as order data, basic user information, etc., and it is difficult to make accurate judgments on the user's all-round needs. However, the user needs in the medical device industry are complex and diverse, including not only structured sales and order data, but also a large amount of unstructured data, such as user consultation records, social media interactions, government bidding information, etc. If these diverse data cannot be fully integrated and analyzed, it will be difficult for companies to accurately grasp market demand, which will lead to a single marketing strategy and poor product promotion results. In addition, the existing systems rely more on static data in the generation of user portraits, making it difficult to dynamically capture changes in user needs, resulting in user portraits lagging behind actual market conditions.

[0004] Secondly, the fragmentation of market information also restricts the decision-making efficiency of enterprises. The medical device industry involves multiple links and complex supply chain networks. From product production, bidding and procurement to sales and logistics, data sources are extensive and highly correlated with each other. Existing systems cannot effectively integrate these cross-domain market data and lack deep insight into the industry's panorama. Most current analysis systems rely on traditional data mining and simple statistical models, making it difficult to discover potential relationships and market opportunities hidden behind the data. In particular, in complex scenarios such as partner selection, policy impact analysis, and market demand forecasting, existing technologies show obvious limitations.

[0005] In addition, the existing big data analysis systems are not very intelligent in providing market decision support. Most systems can only provide simple data reports and trend charts, lack real intelligent decision support, and cannot automatically provide enterprises with the best decision suggestions based on industry knowledge and market dynamics. This decision support system cannot respond to the rapidly changing market environment, making it difficult for enterprises to respond quickly and accurately when facing market competition. Summary of the invention

[0006] The purpose of the present invention is to design an Internet big data analysis method and system based on artificial intelligence, which can realize the comprehensive capture and dynamic update of user needs through the fusion processing of multi-dimensional data and the construction of industry knowledge graphs, and provide intelligent market decision support based on knowledge graphs, thus overcoming the limitations of the existing technology.

[0007] In order to achieve the above object, a first aspect of the present invention provides an Internet big data analysis method based on artificial intelligence, the method comprising:

[0008] Step 1: Collect heterogeneous data from different data sources in the medical device industry, standardize them, extract key features from the standardized data, and fuse them into a unified comprehensive feature vector z f ;

[0009] Step 2: Based on the comprehensive feature vector z f Generate dynamically updated user profiles, including:

[0010] Step 2.1: Based on the comprehensive feature vector z f Use a linear layer plus an activation function to perform demand forecasting. At the same time, add a time-varying weight term λt after the linear layer to adjust the weight of demand forecasting according to the change of time, ensuring that the weight of old data is the smallest while the weight of new data is the largest.

[0011] Step 2.2: Based on the comprehensive feature vector z f Establish an interest preference analysis model and design the policy impact coefficient η p Combine and calculate the user's interest distribution in different categories of products;

[0012] Step 2.3: Based on the comprehensive feature vector z f Design sentiment analysis network based on user feedback, generate sentiment scores, and evaluate user satisfaction with the product;

[0013] Step 2.4: Integrate demand prediction, interest distribution, and sentiment score to generate a comprehensive user portrait. At the same time, design a dynamic update mechanism based on time decay. Whenever new data enters, the portrait will be updated according to the time weight.

[0014] Step 3: Combine the medical device industry with dynamically updated user portraits to build a dynamically updated knowledge graph to keep the market changes up to date, including:

[0015] Step 3.1: Collect market information, which includes characteristic data of medical device products P = {p1, p2, ..., p n}, policy and market information G = {g1,g2,…,g m}, and supply chain partner information S = {s1,s2,…,s k}, build a dynamic knowledge graph based on market information and comprehensive user portraits;

[0016] Step 3.2: Use the graph neural network model to learn the complex relationships between nodes in the dynamic knowledge graph, embed and update the nodes, and capture the structural information in the graph;

[0017] Step 3.3: Design a dynamic update mechanism based on time decay. Whenever new data comes in, the dynamic knowledge graph will automatically update according to the new user behavior and external environment.

[0018] Step 4: Use the knowledge graph to conduct market analysis and provide product demand forecasts and user recommendations, including:

[0019] Through the embedded representation of user nodes, combined with the information of policy nodes and product nodes, the user's future needs and market trends are predicted, and the most suitable products and supply chain partners are recommended to the user based on the predicted results. The user's future needs and market trends are predicted, the most suitable products and the most suitable supply chain partners are generated into comprehensive market strategy recommendations;

[0020] Step 5: Provide personalized marketing plans based on comprehensive market strategy recommendations and optimize in real time.

[0021] Furthermore, the heterogeneous data includes structured data, unstructured data and external data; the standardization process specifically includes the following steps:

[0022] For structured data, we use a custom interpolation algorithm to fill missing values ​​and detect outliers based on historical trends in time series and sales data of similar products. For unstructured data, we use NLP technology to clean text, and then use a custom variant of the pre-trained language model BERT to embed text data into the vector space. For external data, we use NLP technology to extract features and combine them with structured data to obtain the combined structured data.

[0023] Then, the combined structured data and unstructured data are feature extracted and mapped to a unified feature space to obtain a unified feature vector representation z, which is expressed as follows:

[0024] z=W s ·x s +W u ·h i

[0025] Among them, x s Represents a vector in structured data, h i Represents a vector in unstructured data, W s Represents the processed structured data, W u represents unstructured data, W s ∈R p×m and W u ∈R p×d , where R p×m and R p×d is a weight matrix that projects structured and unstructured data into a feature space of the same dimension p, z∈R p is the fused feature vector.

[0026] Furthermore, the key features of the standardized data are extracted and fused into a unified fusion feature vector, specifically including:

[0027] For structured data x s , according to its specific numerical value and time series characteristics, a linear transformation with a regular term is used for mapping, which is expressed as follows:

[0028]

[0029] Among them, z s Represents the mapping vector of structured data, W s ∈R p×m , R p×m It is a linear mapping matrix that maps m-dimensional structured data to p-dimensional unified space. λ is the regularization coefficient, which is used to control the complexity of the model and prevent overfitting.

[0030] For unstructured data i , using the transformation matrix W u Projecting it into p-dimensional space, it is expressed as follows:

[0031] z u =W u ·h i +γ(h i )

[0032] Among them, z u Represents the mapping vector of unstructured data, γ(h i) is a noise suppression term, which is calculated through the attention mechanism or context analysis model and filters irrelevant information in the feedback; W u ∈R p×d , R p×d is the projection matrix, which maps the unstructured text embedding into the p-dimensional unified feature space;

[0033] For external data p , a lag effect control term δ(t) is designed. The lag effect control term δ(t) is related to the policy release time and reflects the gradual impact of the policy on the market, which is expressed as follows:

[0034] z p+1 =W p ·z p δ(t)

[0035] Among them, z p+1 Represents updated external data, W p ∈R p×q , R p×q is the projection matrix;

[0036] The mapping vector z of the structured data s , the mapping vector z of unstructured data u and updated external data z p+1 Fusion into a comprehensive feature vector z f , which is expressed as follows:

[0037] z f =α s z s +α u z u +α p z p

[0038] Among them, α s , α u , α p is the adaptive weight of each modal feature, satisfying α s +α u +α p =1;

[0039] According to the comprehensive feature vector z f Design a dynamic update mechanism to ensure that the feature vector can be adaptively updated over time, as shown below:

[0040]

[0041] Where β is the time decay coefficient, is the current feature vector, The feature vector generated for the most recent data;

[0042] At the same time, in the dynamic update process, the regularization term θ is added to control the smoothness of the feature vector, which is expressed as follows:

[0043]

[0044] Among them, θ is the regularization coefficient, is the Euclidean distance between feature vectors, which is used to control the amplitude of feature vector update to ensure smooth transition.

[0045] Furthermore, in step 2, the demand forecast is based on a dynamic demand forecast model, and the dynamic demand forecast model is constructed as follows:

[0046] Based on the input multidimensional feature vector z f , use a linear layer plus activation function to perform demand forecasting, the formula is as follows:

[0047] y d =σ(W d ·z f +b d )

[0048] Among them, W d is the demand mapping vector, W d ∈R 1×p , R 1×p is a mapping matrix that maps the p-dimensional feature vector to the demand forecast dimension, b d is the bias term, σ is the activation function, y d It indicates the predicted value of the user's future demand, reflecting the possibility of the user purchasing the product in the future;

[0049] Add a regularization term Ω(z f ) balances the impact of historical data and new data, expressed as follows:

[0050]

[0051] Among them, α represents the regularization coefficient, W d represents the mapping matrix, z f represents the comprehensive feature vector;

[0052] The interest preference analysis model is constructed as follows:

[0053] A softmax function is used to map users’ preferences for different product categories into probability distributions, as shown below:

[0054] y i =softmax(W i ·z f +b i)

[0055] Among them, W i represents the interest weight, W i ∈R c×p , R c×p is a weight matrix that transforms the multimodal features z f The interest distribution mapped to c product categories, b i is the bias term;

[0056] The sentiment analysis network is constructed as follows:

[0057] Using a tanh activation function, the output sentiment score range is between [-1, 1]. The specific formula is as follows:

[0058] y f =tanh(W f ·z f +b f )

[0059] Among them, W f represents the sentiment weight, W f ∈R 1×p , R 1×p is the weight matrix, b f is the bias term, y f is the sentiment score, indicating the user's satisfaction with the product; f >0 means the user's feedback on the product is positive, f <0 indicates negative feedback;

[0060] The comprehensive user profile y profile , which is expressed as follows:

[0061]

[0062] Among them, y d represents the user demand forecast, is the interest preference distribution, It is the result of sentiment analysis.

[0063] Furthermore, in the dynamic demand forecasting model, a time-varying weight term λ(t) is introduced to adjust the weight of demand forecasting according to the change of time, which is expressed as follows:

[0064]

[0065] Where λ(t) is the time decay function, represents the updated demand forecast, represents the old demand forecast;

[0066] In the interest preference analysis model, a time decay regularization term ρ is also introduced to balance historical interests and future potential interests, which is expressed as follows:

[0067]

[0068] in, For the latest interest distribution, is the historical interest distribution; at the same time, the policy influence coefficient η is also introduced into the interest preference analysis model p , which is used to enhance the weighted impact of policy changes on interest preferences, is expressed as follows:

[0069]

[0070] Among them, η p reflects the impact of policy changes on interest in different product categories;

[0071] In the sentiment analysis network, the sentiment amplification factor u is also introduced to amplify the sentiment fluctuation, which is expressed as follows:

[0072]

[0073] Among them, μ is the amplification factor. When the user feedback contains extreme emotions, μ will automatically amplify the emotion score to ensure that the user's emotional fluctuations can be fully captured by the system.

[0074] Furthermore, the dynamic knowledge graph structure is as follows:

[0075] The nodes include: user node C, product node P, policy node G, and policy node G;

[0076] The user node C is composed of each user portrait γ profile As the characteristic information of the user node, it includes demand prediction yd, interest preference y i and emotional feedback f ; Through user portrait y profile As feature input, denoted as h C =f(y profile ), where f is the feature embedding function;

[0077] In the product node P, each product node represents a medical device product, denoted as p i , represents the embedding h of product features P =f(p i );

[0078] The policy node G is an industry policy and market information node, including the time of policy release, content and impact scope, etc., represented by g i , represents the embedding h of policy features G=f(g i );

[0079] The supply chain node S represents the node of the supply chain participant, denoted as s i , represents the embedding h of the supply chain participants S =f(s i );

[0080] Relationships include: User-Product Relationship R cp , user-policy relationship R cg and user-supply chain relationship R cs ;

[0081] The user-product relationship R cp The relationship between users and the products they are interested in or have purchased. The edge weight is determined by the user's interest distribution and purchase history;

[0082] The user-policy relationship R cg Indicates the impact of policies on user needs;

[0083] The user-supply chain relationship R cs Represents the interactive relationship between users and supply chain partners, including order delivery history, feedback, and cooperation intensity;

[0084] A dynamic weight mechanism is introduced to construct the adjacency matrix to embed nodes and relationships, which is expressed as follows:

[0085] The user-product adjacency matrix A cp (i, j), is represented as follows:

[0086] A cp (i, j) = α1·sim(y i , p j )+α2·purchase_history(i,j)

[0087] Among them, sim(y i , p j ) represents the similarity between user interests and product features, purchase_history(i, j) represents the historical purchase records of users and products, and α1 and α2 are balance parameters;

[0088] User-policy adjacency matrix A cg (i, j), is represented as follows:

[0089] A cg (i, j) = β1·impact(g j ,y i )+β2·policy_relevance(i, j)

[0090] Among them, impact(g j ,y i ) represents the potential impact of policy on user demand, policy_relevance(i, j) represents the correlation between policy and user demand, β1 and β2 are used to adjust the weights of different factors;

[0091] User-supply chain adjacency matrix A cs (i, j), is represented as follows:

[0092] A cs (i,j)=γ1·order_frequency(i,j)+γ2·logistics_efficiency(i,j)

[0093] Among them, order_frequency(i, j) represents the order frequency between users and suppliers, and logistics_efficiency(i, j) represents the supply chain.

[0094] Furthermore, the use of the graph neural network model to learn the complex relationships between nodes in the dynamic knowledge graph, embed and update the nodes, and capture the structural information in the graph specifically includes:

[0095] In each layer of the graph neural network, the features of the nodes are propagated through the adjacency matrix and interact with neighboring nodes. The features of each node are updated using the following formula:

[0096]

[0097] in, represents the features of node i in the lth layer, N(i) is the set of neighbor nodes of node i, and A ij is the relationship weight between nodes i and j in the adjacency matrix, W (1) is the weight matrix of the lth layer, b (1) is the bias term, σ is the nonlinear activation function;

[0098] At the same time, the relationship regularization term is introduced to control the complexity of the relationship between nodes, which is expressed as follows:

[0099]

[0100] Among them, A R is the adjacency matrix under different relationship types, α is the regularization coefficient, ||·|| F is the Frobenius norm;

[0101] The dynamic update mechanism based on time decay is expressed as follows:

[0102]

[0103] in, is the feature of node i at the current moment, is the feature update brought by the new data, and β is the time decay coefficient, which controls the weight distribution of historical features and new features.

[0104] Furthermore, the embedding representation of the user node is combined with the information of the policy node and the product node to predict the user's future needs and market trends, which is performed as follows:

[0105] Based on the message passing mechanism of graph neural network, the demand prediction of user nodes can be calculated by the following formula:

[0106]

[0107] in, is the prediction result of user demand; W d is the weight matrix that converts user embedding into demand prediction; α g is the policy impact weight, impact(g) is the potential impact of the policy on user demand, and h G is policy embedding; p is the product impact weight, h P Embed for products;

[0108] The method of recommending the most suitable product and the most suitable supply chain partner to the user based on the prediction results is performed as follows:

[0109] Embed h through user nodes C and product nodes are embedded in h P The similarity calculation and the historical purchase relationship between the user and the product are used to generate the recommendation score, which is expressed as follows:

[0110]

[0111] Among them, y rec is the score of product recommendation; W rec is the weight matrix of product recommendation; is the inner product of user embedding and product embedding, which is used to calculate similarity; γ·purchase_history(C, P) is the influence of historical purchase records, where γ is the historical influence coefficient;

[0112] Supply chain partner recommendation is calculated by the embedding similarity of user nodes and supply chain nodes, as well as order frequency and logistics efficiency, as shown below:

[0113]

[0114] Among them, y sc is the supply chain partner recommendation score; W sc is the weight matrix of supply chain recommendations; is the inner product of the embedding of the user and the supply chain node, which is used to calculate the similarity; δ·order_frequency(C, S) represents the order frequency between the user and the supply chain partner, and δ is the order frequency weight; θ·logistics_efficiency(S) is the impact of logistics efficiency on partner recommendation, where θ is the logistics efficiency weight.

[0115] Furthermore, the step 5 specifically includes:

[0116] Step 5.1: Based on demand forecast and product recommendation results, select appropriate marketing channels for different user groups and allocate marketing resources reasonably;

[0117] Step 5.2: Based on user portraits and market feedback data, an adaptive user clustering model is constructed to cluster users, and marketing resources and user clustering results are combined to generate personalized marketing content; wherein, the adaptive user clustering model embeds the user vector h C Perform clustering to generate multiple user groups;

[0118] Step 5.3: Design a real-time feedback mechanism to adjust the current marketing strategy and channel resource allocation by comparing the actual feedback data of personalized marketing content with the expected results. At the same time, design a step control mechanism to ensure that the strategy optimization process can converge and avoid strategy fluctuations caused by excessive adjustments.

[0119] In a second aspect of the present invention, an artificial intelligence-based Internet big data analysis system is provided, the system comprising:

[0120] The heterogeneous data collection module is used to collect heterogeneous data from different data sources in the medical device industry, standardize them, extract key features from the standardized data, and fuse them into a unified comprehensive feature vector z f ;

[0121] The image building module is used to build the image according to the comprehensive feature vector z f Generate dynamically updated user profiles, including:

[0122] Step 2.1: Based on the comprehensive feature vector z f A linear layer plus an activation function is used to predict demand. At the same time, a time-varying weight term λ(t) is added after the linear layer to adjust the weight of demand prediction according to the change of time, ensuring that the weight of old data is the smallest while the weight of new data is the largest.

[0123] Step 2.2: Based on the comprehensive feature vector zf Establish an interest preference analysis model and design the policy impact coefficient η p Combine and calculate the user's interest distribution in different categories of products;

[0124] Step 2.3: Based on the comprehensive feature vector z f Design sentiment analysis network based on user feedback, generate sentiment scores, and evaluate user satisfaction with the product;

[0125] Step 2.4: Integrate demand prediction, interest distribution, and sentiment score to generate a comprehensive user portrait. At the same time, design a dynamic update mechanism based on time decay. Whenever new data enters, the portrait will be updated according to the time weight.

[0126] The dynamic knowledge graph construction module is used to build a dynamically updated dynamic knowledge graph based on the medical device industry and dynamically updated user portraits, so that market changes are kept up to date. Specifically, it includes:

[0127] Step 3.1: Collect market information, which includes characteristic data of medical device products P = {p1, p2, ..., p n}, policy and market information G = {g1,g2,…,g m}, and supply chain partner information S = {s1,s2,…,s k}, build a dynamic knowledge graph based on market information and comprehensive user portraits;

[0128] Step 3.2: Use the graph neural network model to learn the complex relationships between nodes in the dynamic knowledge graph, embed and update the nodes, and capture the structural information in the graph;

[0129] Step 3.3: Design a dynamic update mechanism based on time decay. Whenever new data comes in, the dynamic knowledge graph will automatically update according to the new user behavior and external environment.

[0130] The dynamic knowledge graph analysis module is used to use the knowledge graph to conduct market analysis, provide product demand forecasts and user recommendations, including:

[0131] Through the embedded representation of user nodes, combined with the information of policy nodes and product nodes, the user's future needs and market trends are predicted, and the most suitable products and supply chain partners are recommended to the user based on the predicted results. The user's future needs and market trends are predicted, the most suitable products and the most suitable supply chain partners are generated into comprehensive market strategy recommendations;

[0132] The real-time optimization module is used to provide personalized marketing plans and optimize in real time based on comprehensive market strategy recommendations.

[0133] The beneficial technical effects of the present invention are at least as follows:

[0134] First of all, the present invention breaks through the limitations of existing big data analysis systems in data processing dimensions through multi-dimensional data fusion technology. The system can integrate structured and unstructured data in the medical device industry, including user purchasing behavior, online consultation records, government bidding information, industry policies, etc., and use deep learning and natural language processing technology to process and fuse multimodal data. Compared with the shortcomings of traditional technology that only relies on a single structured data source, the present invention can generate a more accurate and comprehensive user portrait. More importantly, the user portrait is dynamically updated, and can track user behavior changes in real time to ensure that the portrait reflects the user's latest needs, significantly improving the accuracy of the marketing strategy.

[0135] Secondly, the present invention solves the problem that existing systems cannot integrate cross-domain market information by constructing a knowledge graph for the medical device industry. The knowledge graph is based on different entities in the industry (such as products, users, suppliers) and their relationships, forming a networked association between data, which can help the system discover market opportunities and industry rules hidden behind the data. By analyzing these relationships, the system can identify potential changes in market demand, the impact of policies on the market, and find partners suitable for the company. Compared with traditional systems that can only provide data reports and trend charts, the knowledge graph of the present invention can provide a deeper level of industry analysis, which is particularly suitable for complex decision-making scenarios such as market trend forecasting and partner recommendations.

[0136] Finally, the present invention provides intelligent market decision support functions through in-depth analysis of knowledge graphs, overcoming the problem of insufficient intelligence level of existing systems. The system can not only generate conventional market analysis reports, but also automatically identify potential opportunities in the market based on knowledge graphs, predict product demand, and even provide personalized marketing strategy suggestions. Enterprises can use this system to intelligently optimize marketing plans, select the best partners, and conduct market layout, significantly improving market response speed and decision-making accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0137] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0138] Figure 1 This is a flow chart of the Internet big data analysis method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0139] Embodiments of the present invention are described in detail below, examples of which 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 only used to explain the present invention, and cannot be understood as limiting the present invention.

[0140] In one or more embodiments, Figure 1 As shown, the present invention discloses an Internet big data analysis method based on artificial intelligence, which includes steps S1-S5:

[0141] Step 1: Collect heterogeneous data from different data sources in the medical device industry, standardize them, extract key features from the standardized data, and fuse them into a unified comprehensive feature vector z f .

[0142] Specifically, first, the present invention obtains multi-dimensional data of the medical device industry from multiple sources, including but not limited to the following categories:

[0143] D s : Structured data, such as sales records, order information, user purchase behavior, etc. s Each record can be represented as a vector x s =[x s1 , x s2 , ..., x sm ], where m is the dimension of the structured data.

[0144] D u : Unstructured data, mainly text data from users' consultation records, feedback information, and social media interactions. The present invention represents unstructured data as a collection of text D u ={t1, t2, ..., t n}, where t i For a single text.

[0145] D e : External data, such as government bidding information, industry policy changes, etc. These data are public, usually in the form of reports and announcements, and can be represented by text data.

[0146] All data sources D = D s ∪D u ∪D e All of them need to be standardized to ensure the consistency and operability of the data. The multimodality of the input data is a major difficulty in this patent.

[0147] Furthermore, for structured data D s, missing value filling and outlier detection are required. The present invention adopts a custom interpolation algorithm to fill missing values. This method performs dynamic interpolation based on the historical trends in the time series and the sales data of similar products. The specific calculation formula is as follows:

[0148]

[0149] in, For the filled data points, is the historical data of similar product i on feature j, w i is the weight based on historical similarity. This interpolation method can dynamically adjust the weight to ensure adaptability to the similarities between different products, thereby handling the accuracy of missing data.

[0150] For unstructured data D u , the present invention uses NLP technology to clean text, including removing stop words, punctuation marks and redundant information. Afterwards, the text data is embedded into the vector space using a custom variant of the pre-trained language model BERT. Specifically, each text segment t i After the BERT model transformation, the vectorized representation h is obtained. i ,Right now:

[0151] h i =BERT(t i )

[0152] where h i ∈R d is the feature representation of the text, and d is the dimension of the embedding space. The embedded text data will be used as part of the subsequent data fusion.

[0153] For external data D e The present invention extracts the features of key policy changes or bidding information through NLP technology and combines them with structured data. For example, keywords related to market impact are extracted from government policy texts and converted into actionable indicator information.

[0154] Furthermore, after the data is cleaned and standardized, different types of data are then subjected to feature extraction and mapped into a unified feature space. In order to integrate structured and unstructured data, the present invention designs a self-supervised learning model that can extract and fuse features from multimodal data.

[0155] First, for structured data D s , the present invention directly converts the feature x s As input. For unstructured text data, the embedded representation h has been obtained through the BERT model i Next, the present invention defines a fusion function f(xs ,h i ), combining the data features of these two different modalities into a unified representation. The specific fusion calculation is as follows:

[0156] z=W s ·x s +W u ·h i

[0157] Among them, W s ∈R p×m and W u ∈R p×d is a weight matrix that projects structured and unstructured data into a feature space of the same dimension p, z∈R p is the fused feature vector. This fusion process is designed for the special needs in the field of the present invention, because the needs of users in the medical device industry not only depend on purchasing behavior, but also on unstructured feedback and policy information.

[0158] Finally, after all types of data are processed as above, a unified feature vector representation z is obtained. This feature vector not only includes the user's purchase behavior, but also combines external policy information and user feedback. Each feature vector z is an important input for the subsequent generation of user portraits and knowledge graphs.

[0159] Since the behavior and needs of each user are time-sensitive, the feature vector z needs to be updated dynamically. For example, when new sales data or feedback information enters the system, the feature vector will be recalculated in real time and sent to the subsequent modules. This dynamic update mechanism is one of the keys to the design of this system for the medical device field, which can ensure that the user portrait remains accurate at all times.

[0160] Specifically, the present invention designs a multimodal feature mapping network to map data of different modalities into a unified feature space. Aiming at the unique properties of each type of data in the medical device industry, the present invention designs different mapping rules for structured data, unstructured data and policy data.

[0161] First, for the structured data x s , since they are specific numerical and time series features, the present invention uses a linear transformation with a regular term for mapping, especially considering the overfitting phenomenon in historical data, so a penalty term λ is added in the transformation process to control overly complex mapping:

[0162]

[0163] Among them, W s ∈R p×mis a linear mapping matrix that maps m-dimensional structured data to p-dimensional uniform space, and λ is the regularization term coefficient, which is used to control model complexity and prevent overfitting. This is particularly important for the medical device industry, because order and sales data often have large fluctuations and must be controlled by regularization terms.

[0164] For unstructured text embedding h i , the present invention adopts another transformation matrix W u It is projected into the p-dimensional space. At the same time, text data often contains potential noise (such as irrelevant user feedback or emotions), so the present invention designs an adaptive noise reduction term γ(h i ), noise suppression is performed based on the contextual importance weight of the text:

[0165] z u =W u ·h i +γ(h i )

[0166] Among them, γ(h i ) is a noise suppression term, which is calculated through the attention mechanism or context analysis model to filter irrelevant information in the feedback. u ∈R p×d is the projection matrix, which maps the unstructured text embedding into a p-dimensional unified feature space.

[0167] For external policy data p ,Since the policy has a lag and indirect effect on market demand, this paper designs a lag effect control term δ(t), which is related to the policy release time and reflects the gradual impact of the policy on the market:

[0168] z p =W p ·z p δ(t)

[0169] Among them, W p ∈R p×q is the projection matrix, δ(t) is the time lag function, and the lag effect gradually weakens as time t passes. This formula fully considers the timeliness of policy data and ensures that the impact of policies on user demand and purchasing behavior is fully captured.

[0170] Furthermore, in order to unify the data of three different modalities into a comprehensive feature vector, the present invention adopts an adaptive weighted fusion strategy. The present invention designs an adaptive weight generation network, which dynamically assigns weights according to the data characteristics of each modality to ensure that the more important modality obtains a higher weight. The fusion formula is as follows:

[0171] z f =αs z s +α u z u +α p z p

[0172] Among them, α s , α u , α p is the adaptive weight of each modal feature, satisfying α s +α u +α p = 1, which are generated by network adaptiveness. These weights are not only related to the modality, but also to factors such as data timeliness and data quality. The result of this step is the fused unified feature vector z f ∈R p The specially designed adaptive weights ensure that sudden changes in medical device industry data (such as new policies, major feedback, etc.) can be captured in a timely manner, avoiding the lag problem caused by fixed weights in traditional weighting methods.

[0173] Furthermore, the eigenvector z f After generation, since user behavior and market dynamics will continue to change, the present invention must design a dynamic update mechanism to ensure that the feature vector can be adaptively updated over time. The present invention introduces a time-weighted update algorithm:

[0174]

[0175] Where β is the time decay coefficient, is the current feature vector, The feature vector generated for the latest data. This weighted update mechanism ensures that new data has a timely impact on user profiles while retaining the long-term memory effect of some historical data.

[0176] In addition, in the dynamic update mechanism, the present invention also adds a regularization term θ to control the smoothness of the feature vector to avoid large fluctuations of the feature vector caused by excessive noise:

[0177]

[0178] Among them, θ is the regularization coefficient, is the Euclidean distance between feature vectors, which is used to control the amplitude of feature vector update to ensure smooth transition.

[0179] In summary, the process of feature extraction and multimodal fusion is divided into four steps: 1) input data initialization; 2) cross-modal feature transformation; 3) feature fusion; 4) dynamic update of feature vectors. Through this systematic feature extraction solution, the problem of complex, changeable and dynamic update of multimodal data in the medical device industry is solved, and highly accurate input is provided for subsequent user profiling and intelligent market decision-making.

[0180] Step 2: Based on the comprehensive feature vector z f Generate dynamically updated user profiles.

[0181] Specifically, demand forecasting is one of the core tasks of user portrait generation. Especially for the medical device industry, user demand is often affected by policy changes, seasonal fluctuations, and the user's own historical behavior. In order to capture these complex influencing factors, this paper designs an innovative dynamic demand forecasting model based on z f Predict the user's future needs. Based on the input multidimensional feature vector z f , the present invention uses a linear layer plus an activation function to perform demand forecasting, and the formula is as follows:

[0182] y d =σ(W d ·z f +b d )

[0183] Among them, W d ∈R 1×p is a mapping matrix that maps the p-dimensional feature vector to the demand forecast dimension, b d is the bias term, σ is the activation function, and ReLU or Sigmoid is usually used to limit the output range. d It represents the predicted value of the user's future demand and reflects the possibility of the user purchasing the product in the future.

[0184] In order to consider the impact of external policies or market fluctuations on demand, the present invention introduces a time-varying weight term λ(t), which is used to adjust the weight of demand forecast according to the change of time, ensuring that the weight of old data gradually decreases and the weight of new data gradually increases:

[0185]

[0186] Among them, λ(t) is the time decay function, which can be adjusted dynamically through exponential decay or based on historical data. At this time, the variable enables the system to adapt to changes in the market or policy and adjust the results of demand forecasting in a timely manner.

[0187] Furthermore, in the process of demand forecasting, the user demand in the medical device industry is usually affected by some unforeseen fluctuations. In order to avoid the model over-relying on historical data and ignoring potential new demand changes, the present invention designs an innovative regularization term Ω(z f ), used to balance the impact of historical data and new data:

[0188]

[0189] This item controls the overfitting problem of model parameters in demand forecasting and ensures that the feature map has a certain smoothness and robustness, so that the model can adapt to fluctuations in demand.

[0190] Furthermore, in addition to demand forecasting, user interest preferences for different products are also an important component of accurate user portraits. f , establish an interest preference analysis model to calculate the user's interest distribution for different categories of products. The present invention uses a softmax function to map the user's preference for different product categories into a probability distribution. The specific formula is as follows:

[0191] y i =softmax(W i ·z f +b i )

[0192] Among them, W i ∈R c×p is a weight matrix that transforms the multimodal features z f The interest distribution mapped to c product categories, b i is the bias term. Through the softmax function, the model outputs a normalized interest distribution, y i Indicates the user's interest level in each product category.

[0193] In order to prevent interest preference analysis from being overly dependent on the user's historical purchase behavior, resulting in a single recommendation result, the present invention introduces a time decay regularization term ρ to balance historical interests and future potential interests:

[0194]

[0195] in, For the latest interest distribution, is the historical interest distribution. Through this regularization term, the present invention can dynamically adjust in the recommendation, prevent the user profile from being overly dependent on historical data, and give new products a higher recommendation weight.

[0196] Furthermore, since the product update and iteration speed of the medical device industry is slow, but the policies and demands change rapidly, a policy impact coefficient η is also added to the model. p , specifically used to increase the weight of policy changes on interest preferences:

[0197]

[0198] Among them, η p It reflects the impact of policy changes on the interest in different product categories, ensuring that after certain policies are introduced, related product categories can receive a higher recommendation weight.

[0199] Furthermore, in the medical device industry, user feedback and emotional tendencies will have an important impact on future purchasing behavior and demand. The present invention designs a sentiment analysis network that combines text feedback data and multimodal fusion features. f , generate a sentiment score y f , used to evaluate user satisfaction with the product. The sentiment score model uses a tanh activation function, and the output sentiment score range is between [-1, 1]. The specific formula is as follows:

[0200] y f =tanh(W f ·z f +b f )

[0201] in, is the weight matrix, b f is the bias term, y f is the sentiment score, indicating the user’s satisfaction with the product. f >0 means that users have positive feedback on the product, f <0 indicates negative feedback.

[0202] Furthermore, in response to the situation where the user's emotions change greatly (for example, strong dissatisfaction or extreme satisfaction with certain products), the present invention introduces an emotion amplification factor μ to amplify the emotion fluctuations:

[0203]

[0204] Among them, μ is the amplification factor. When the user feedback contains extreme emotions, μ will automatically amplify the emotion score to ensure that the user's emotional fluctuations can be fully captured by the system.

[0205] Furthermore, after demand prediction, interest preference analysis and sentiment analysis, the present invention finally generates a comprehensive user profile y profile This portrait can fully reflect the user's potential needs, interest distribution and emotional tendencies. The specific formula is as follows:

[0206]

[0207] Among them, y d represents the user demand forecast, is the interest preference distribution, This is the result of sentiment analysis. This portrait will be used to formulate precise marketing strategies and personalized product recommendations in subsequent steps.

[0208] Furthermore, in order to ensure the timeliness of user portraits, the present invention designs a dynamic update mechanism based on time decay. Whenever new data enters the system, the portrait will be updated according to the time weight:

[0209]

[0210] Among them, β is the time decay coefficient, which controls the weight distribution of old and new portraits. This ensures that the portrait retains historical information while capturing new needs, and generates a dynamic user portrait with timeliness.

[0211] Step 3: Combine the medical device industry with dynamically updated user portraits to build a dynamically updated knowledge graph to keep market changes up to date.

[0212] Specifically, by generating a user portrait The portrait includes the results of user demand prediction, interest preference distribution and sentiment analysis. The present invention also obtains the characteristic data P of the medical device product in the previous step = {p1, p2, ..., p n}, policy and market information G = {g1, g2, ..., g m}, and supply chain partner information S = {s1, s2, ..., s k}. These data will be used to build a dynamic knowledge graph.

[0213] The purpose of this step is to integrate these structured and unstructured data from different sources into a knowledge graph KG that can describe different entities and their relationships in the medical device industry, such as users, products, policies, and supply chains, and use this graph to discover potential market opportunities, optimize supply chain partnerships, and better serve user needs. The knowledge graph will play a key role in the subsequent intelligent decision-making process.

[0214] Furthermore, the construction of the knowledge graph first requires defining each type of entity in the data as a node of the graph, and the relationships between different entities as edges of the graph.

[0215] Node definition:

[0216] User node C: each user's profile y profileAs the characteristic information of the user node, it contains the demand prediction y d 、Interest Preference i and emotional feedback f .

[0217] Product node P: Each product node represents a medical device product. The characteristics of the node include price, performance, type, etc., expressed as p i .

[0218] Policy node G: Industry policy and market information node, including the time of policy release, content and impact scope, etc., represented by g i .

[0219] Supply chain node S: Nodes representing supply chain participants such as manufacturers, distributors, and logistics companies, denoted as s i .

[0220] Relationship definition:

[0221] The relationship between different nodes is represented by edges:

[0222] User-Product Relationship R cp : Defines the relationship between users and products they are interested in or have purchased. Edge weights are determined by user interest distribution and purchase history.

[0223] User-policy relationship R cg : Indicates the impact of policies on user demand. Policies may affect user demand through price fluctuations, product compliance, and supply chain changes.

[0224] User-supply chain relationship R cs : Represents the interactive relationship between users and supply chain partners (such as manufacturers or logistics providers), including order delivery history, feedback, and cooperation intensity.

[0225] Furthermore, in order to effectively model nodes and their relationships, especially the interaction between users and external environments (such as policies and supply chains), the present invention introduces an adaptive graph embedding model to map different entities and their relationships into a common feature space.

[0226] Node embedding:

[0227] The initial features of each node come from the previously generated data and are expressed as:

[0228] User node embedding h C :Through user portrait y profile As feature input, denoted as h C =f(y profile ), where f is the feature embedding function.

[0229] Product node embedding hP : Embedding h representing product features P =f(p i ).

[0230] Policy node embedding h G : Embedding h representing policy features G =f(g i ).

[0231] Supply chain node embedding h S : represents the embedding h of the supply chain participants S =f(s i ).

[0232] Adjacency matrix construction: The relationship between different entities is represented by the adjacency matrix. The present invention introduces a dynamic weight mechanism to construct the adjacency matrix. These weights change with time or data update. The definition is as follows:

[0233] User-product adjacency matrix A cp :

[0234] A cp (i, j) = α1·sim(y i , p j )+α2·purchase_history(i,j)

[0235] Among them, sim(y i , p j ) represents the similarity between user interests and product features, purchase_history(i, j) represents the historical purchase records of users and products, and α1 and α2 are balance parameters.

[0236] User-policy adjacency matrix A cg :

[0237] A cg (i, j) = β1·impact(g j ,y i )+β2·policy_relevance(i, j)

[0238] Among them, impact(g j ,y i ) represents the potential impact of policy on user demand, and policy_relevance(i, j) represents the relevance between policy and user demand. β1 and β2 are used to adjust the weights of different factors.

[0239] User-supply chain adjacency matrix A cs :

[0240] A cs(i,j)=γ1·order_frequency(i,j)+γ2·logistics_efficiency(i,j)

[0241] Among them, order_frequency(i, j) represents the order frequency between users and suppliers, logistics_efficiency(i, j) represents the logistics efficiency of the supply chain, and γ1 and γ2 control the weights of each factor.

[0242] Furthermore, in order to learn the complex relationships between nodes in the graph and provide support for subsequent intelligent decision-making, the present invention uses a graph neural network to embed and update the nodes to capture the structural information in the graph.

[0243] Furthermore, in each layer of the graph neural network, the features of the nodes are propagated through the adjacency matrix and interact with neighboring nodes. The features of each node are updated using the following formula:

[0244]

[0245] in, represents the feature of node i in layer l, is the set of neighbor nodes of node i, A ij is the relationship weight between nodes i and j in the adjacency matrix, W (l) is the weight matrix of the lth layer, b (l) is the bias term, and σ is a nonlinear activation function (such as ReLU). This formula describes how the user node obtains information from the product, policy, and supply chain nodes and updates the feature representation through multi-layer propagation. This process can effectively capture the potential associations between nodes, such as the indirect impact of a policy on user demand or the impact of supply chain logistics efficiency on user satisfaction.

[0246] Furthermore, in order to ensure that the embedding representation of the graph neural network is not overly complex while retaining important relationships in the graph structure, the present invention introduces a relation regularization term to control the complexity of the relationship between nodes:

[0247]

[0248] Among them, A R is the adjacency matrix under different relationship types, α is the regularization coefficient, ||·|| F is the Frobenius norm. Through this regularization term, the present invention can prevent the graph structure from being too complex, thereby maintaining the sparsity of relations in the knowledge graph, making it easier to explain and apply.

[0249] Furthermore, in the medical device industry, factors such as user demand, market policies, and supply chain efficiency are constantly changing. Therefore, the update mechanism of the knowledge graph must be dynamically adaptable.

[0250] Furthermore, the present invention designs a dynamic update mechanism based on time decay. Whenever new data enters the system, the knowledge graph will automatically update according to the new user behavior and external environment:

[0251]

[0252] in, is the feature of node i at the current moment, is the feature update brought by new data, and β is the time decay coefficient, which controls the weight distribution of historical features and new features. This mechanism ensures that the knowledge graph can change with market dynamics, stay up to date, and provide real-time support for intelligent decision-making.

[0253] Step 4: Use the knowledge graph to conduct market analysis and provide product demand forecasts and user recommendations.

[0254] Specifically, the present invention firstly represents h by embedding the user node C , combined with the policy node h G and product node h P information to predict future user needs and market trends.

[0255] Based on the message passing mechanism of graph neural network, the demand prediction of user nodes can be calculated by the following formula:

[0256]

[0257] in, is the prediction result of user demand; W d is the weight matrix that converts user embedding into demand prediction; α g is the policy impact weight, impact(g) is the potential impact of the policy on user demand, and h G is policy embedding; p is the product impact weight, h P is product embedding. This formula combines the influence of user embedding, policy and product to predict future changes in user demand. In order to capture the dynamic changes in the market, the influence weights of policy nodes and product nodes α g and β p It can be learned automatically through the model training process.

[0258] Furthermore, based on the demand forecast, the system further recommends the most suitable products to the user. The present invention constructs an adaptive product recommendation model based on the interest distribution of the user portrait, the impact of the policy on the market and the user's historical purchase record.

[0259] Embed h through user nodes C and product nodes are embedded in h P The similarity calculation and the historical purchase relationship between users and products are used to generate recommendation scores:

[0260]

[0261] Among them, y rec is the score of product recommendation; W rec is the weight matrix of product recommendation; is the inner product of user embedding and product embedding, used to calculate similarity; γ·purchase_history(C, P) is the impact of historical purchase records, where γ is the historical impact coefficient. Product recommendation scores combine the similarity between users and products, historical behaviors, and policy impacts to achieve precise personalized recommendations. The softmax function ensures that the recommended products have a normalized score, and the final product list provided to the user will be sorted by score.

[0262] Furthermore, in addition to forecasting demand and product recommendations, the system also needs to recommend the most suitable supply chain partners for enterprises to optimize supply chain efficiency, reduce costs, and increase order fulfillment speed. S The embedding representation is used to make recommendations by analyzing user needs, product information, and supply chain history data. Supply chain partner recommendations are calculated by the embedding similarity between user nodes and supply chain nodes, as well as order frequency and logistics efficiency:

[0263]

[0264] Among them, y sc is the supply chain partner recommendation score; W sc is the weight matrix of supply chain recommendations; is the inner product of the embedding of the user and the supply chain node, which is used to calculate the similarity; δ·order_frequency(C, S) represents the order frequency between the user and the supply chain partner, and δ is the order frequency weight; θ·logistics_efficiency(S) is the impact of logistics efficiency on partner recommendation, where θ is the logistics efficiency weight.

[0265] Through this formula, the system can integrate user needs, product features, order history and supply chain efficiency to recommend the most suitable partners for enterprises and help them optimize operations.

[0266] Furthermore, finally, based on the results of demand forecasting, product recommendations, and supply chain partner recommendations, the system generates comprehensive market strategy recommendations. Strategies include product promotion plans, target user group selection, partner optimization, etc. All strategies can respond to changes in market and user needs in a timely manner through dynamic updates of the knowledge graph and adaptive adjustments of the prediction model.

[0267] The system also provides enterprises with visual decision-making outputs, including market trend charts, user demand forecast changes, supply chain cooperation suggestions, etc. This information can be used for business analysis by decision makers to help them better plan resources and market promotion.

[0268] Step 5: Provide personalized marketing plans based on comprehensive market strategy recommendations and optimize in real time.

[0269] Specifically, it is necessary to first select appropriate marketing channels for different user groups and reasonably allocate marketing resources. The selection and allocation of marketing channels should be based on the user's demand intensity, product recommendation priority and historical marketing results. The present invention designs an innovative channel weight allocation model, which is based on user demand prediction and recommendation score, and combines the historical performance of the channel to dynamically adjust the priority of each channel. The specific weight allocation formula is as follows:

[0270]

[0271] Among them, w c Represents the weight of marketing channel c, controlling the resource allocation ratio of each channel; Represents the predicted value of user demand, which is used to measure the user's future potential demand; rec is the product recommendation score, which indicates the priority of recommending products to users; conv (c) is the historical conversion rate of marketing channel c, indicating the performance of the channel in past marketing activities; λ1, λ2, and λ3 are adjustment coefficients used to balance the impact of demand, recommendation, and historical performance.

[0272] This formula innovatively combines user demand forecasts, product recommendation priorities, and historical conversion rates to dynamically optimize channel weights while ensuring rational allocation of resources, achieving more efficient resource utilization.

[0273] Furthermore, in the medical device industry, users have different needs and preferences, so it is crucial to achieve precision marketing through user grouping. Based on the user portrait and market feedback data in step 4, the present invention groups users through an adaptive user grouping model. Each group can obtain different marketing strategies and personalized content recommendations to ensure that the marketing content matches the user's interests.

[0274] Furthermore, user clustering is performed by embedding the user vector h C Clustering is performed to generate multiple user groups. The personalized marketing content for each group is generated by weighting dimensions such as user interest distribution and demand prediction. The formula is as follows:

[0275]

[0276] Among them, y content (g) represents the personalized marketing content of user group g; W content is the weight matrix of content generation; aggregate(h C , g) represents the feature aggregation operation of users in user group g; α is the weight of recommendation priority, y rec (i) is the product recommendation score; β is the demand prediction weight, The demand forecast result for this user group. This formula aggregates the needs, product recommendations and user characteristics of user groups to generate personalized marketing content, ensuring that the content is highly matched with user needs, thereby improving marketing conversion rates.

[0277] Furthermore, feedback data (such as click-through rate, conversion rate, etc.) in marketing activities provides an important basis for strategy optimization. The present invention designs a real-time feedback mechanism, which continuously adjusts the current marketing strategy and channel resource allocation by comparing the actual feedback data with the expected effect.

[0278] Furthermore, the channel weights are adjusted through real-time feedback. If the actual performance of a channel is better than expected, the resource investment will be increased; otherwise, the investment will be reduced. The specific formula is as follows:

[0279]

[0280] in, is the updated channel weight; η is the feedback adjustment coefficient, which controls the weight adjustment range; actual_rate(c) is the actual conversion rate of channel c; expected_rate(c) is the expected conversion rate of channel c. This feedback adjustment formula can adjust resource allocation in time according to market feedback, so that marketing strategies can adapt to the dynamic market environment and maintain high efficiency.

[0281] Furthermore, in the medical device industry, factors such as market policies and user needs are changing dynamically, so the optimization of marketing strategies requires adaptive adjustment. The present invention designs a step control mechanism to ensure that the strategy optimization process can gradually converge and avoid excessive adjustments leading to strategy fluctuations.

[0282] Furthermore, after each marketing cycle, the step length η gradually decays to ensure that the strategy gradually converges. The formula is:

[0283] η new =η old γ

[0284] Among them, η new is the new step size; η old is the current step size; γ is the step size attenuation coefficient, which is usually between (0, 1). Through step size control, the system can dynamically optimize the marketing strategy, ensuring that the strategy gradually converges while responding to market changes and avoiding large fluctuations.

[0285] Finally, after multiple strategy adjustments and optimizations, the system will generate the final marketing strategy output, including:

[0286] Channel resource allocation suggestions: resource allocation ratios for different channels;

[0287] User segmentation and personalized content recommendations: personalized marketing strategies for different user groups;

[0288] Long-term strategy optimization path: Based on historical data and market feedback, provide optimization suggestions for future marketing strategies.

[0289] The system continuously acquires new market data and performs loop feedback optimization to ensure that marketing strategies can continue to improve in the long term.

[0290] In this step, by integrating market decision data, intelligent marketing strategy optimization realizes a complete optimization process of marketing channel selection and resource allocation, user grouping and personalized content generation, real-time feedback and dynamic adjustment. Through innovative channel weight allocation model, personalized content generation formula and adaptive step control mechanism, the system can dynamically respond to market changes, ensure continuous optimization and gradual convergence of marketing strategies, and thus improve overall marketing efficiency.

[0291] A second embodiment of the present invention provides an Internet big data analysis system based on artificial intelligence, the system comprising:

[0292] The heterogeneous data collection module 101 is used to collect heterogeneous data from different data sources in the medical device industry, perform standardization processing, extract key features from the standardized data, and fuse them into a unified comprehensive feature vector z f ;

[0293] The portrait construction module 102 is used to construct the image according to the comprehensive feature vector z f Generate dynamically updated user profiles, including:

[0294] Step 2.1: Based on the comprehensive feature vector z fA linear layer plus an activation function is used to predict demand. At the same time, a time-varying weight term λ(t) is added after the linear layer to adjust the weight of demand prediction according to the change of time, ensuring that the weight of old data is the smallest while the weight of new data is the largest.

[0295] Step 2.2: Based on the comprehensive feature vector z f Establish an interest preference analysis model and design the policy impact coefficient η p Combine and calculate the user's interest distribution in different categories of products;

[0296] Step 2.3: Based on the comprehensive feature vector z f Design sentiment analysis network based on user feedback, generate sentiment scores, and evaluate user satisfaction with the product;

[0297] Step 2.4: Integrate demand prediction, interest distribution, and sentiment score to generate a comprehensive user portrait. At the same time, design a dynamic update mechanism based on time decay. Whenever new data enters, the portrait will be updated according to the time weight.

[0298] The dynamic knowledge graph construction module 103 is used to construct a dynamically updated dynamic knowledge graph by combining the medical device industry with the dynamically updated user portraits, so that the market changes are kept up to date, specifically including:

[0299] Step 3.1: Collect market information, which includes characteristic data of medical device products P = {p1, p2, ..., p n}, policy and market information G = {g1,g2,…,g m}, and supply chain partner information S = {s1,s2,…,s k}, build a dynamic knowledge graph based on market information and comprehensive user portraits;

[0300] Step 3.2: Use the graph neural network model to learn the complex relationships between nodes in the dynamic knowledge graph, embed and update the nodes, and capture the structural information in the graph;

[0301] Step 3.3: Design a dynamic update mechanism based on time decay. Whenever new data comes in, the dynamic knowledge graph will automatically update according to the new user behavior and external environment.

[0302] The dynamic knowledge graph analysis module 104 is used to use the knowledge graph to perform market analysis, provide product demand forecasts and user recommendations, and specifically includes:

[0303] Through the embedded representation of user nodes, combined with the information of policy nodes and product nodes, the user's future needs and market trends are predicted, and the most suitable products and supply chain partners are recommended to the user based on the predicted results. The user's future needs and market trends are predicted, the most suitable products and the most suitable supply chain partners are generated into comprehensive market strategy recommendations;

[0304] The real-time optimization module 105 is used to provide personalized marketing plans based on comprehensive market strategy recommendations and optimize them in real time.

[0305] In summary, the present invention proposes an Internet big data analysis method and system based on artificial intelligence. Through the fusion processing of multi-dimensional data and the construction of industry knowledge graphs, it can achieve comprehensive capture and dynamic updating of user needs, and provide intelligent market decision support based on knowledge graphs, thus overcoming the limitations of existing technologies.

[0306] The above disclosures are only some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention are still within the scope of the invention.

Claims

1. An Internet big data analysis method based on artificial intelligence, characterized in that: The method comprises: Step 1: Collect heterogeneous data from different data sources in the medical device industry, standardize them, extract key features from the standardized data, and fuse them into a unified comprehensive feature vector z f ; Step 2: Based on the comprehensive feature vector z f Generate dynamically updated user profiles, including: Step 2.1: Based on the comprehensive feature vector z f A linear layer plus an activation function is used to predict demand. At the same time, a time-varying weight term λ(t) is added after the linear layer to adjust the weight of demand prediction according to the change of time, ensuring that the weight of old data is the smallest while the weight of new data is the largest. Step 2.2: Based on the comprehensive feature vector z f Establish an interest preference analysis model and design the policy impact coefficient η p Combine and calculate the user's interest distribution in different categories of products; Step 2.3: Based on the comprehensive feature vector z f Design sentiment analysis network based on user feedback, generate sentiment scores, and evaluate user satisfaction with the product; Step 2.4: Integrate demand prediction, interest distribution, and sentiment score to generate a comprehensive user portrait. At the same time, design a dynamic update mechanism based on time decay. Whenever new data enters, the portrait will be updated according to the time weight. Step 3: Combine the medical device industry with dynamically updated user portraits to build a dynamically updated knowledge graph to keep the market changes up to date, including: Step 3.1: Collect market information, which includes characteristic data of medical device products P = {p1, p2, ., p n }, policy and market information G = {g1, g2, ..., gm}, and supply chain partner information S = {s1, s2, ..., s k }, build a dynamic knowledge graph based on market information and comprehensive user portraits; Step 3.2: Use the graph neural network model to learn the complex relationships between nodes in the dynamic knowledge graph, embed and update the nodes, and capture the structural information in the graph; Step 3.3: Design a dynamic update mechanism based on time decay. Whenever new data comes in, the dynamic knowledge graph will automatically update according to the new user behavior and external environment. Step 4: Use the knowledge graph to conduct market analysis and provide product demand forecasts and user recommendations, including: Through the embedded representation of user nodes, combined with the information of policy nodes and product nodes, the user's future needs and market trends are predicted, and the most suitable products and supply chain partners are recommended to the user based on the predicted results. The user's future needs and market trends are predicted, the most suitable products and the most suitable supply chain partners are generated into comprehensive market strategy recommendations; Step 5: Provide personalized marketing plans based on comprehensive market strategy recommendations and optimize in real time.

2. The method for analyzing Internet big data based on artificial intelligence according to claim 1, characterized in that: The heterogeneous data includes structured data, unstructured data and external data; the standardization process specifically includes the following steps: For structured data, we use a custom interpolation algorithm to fill missing values ​​and detect outliers based on historical trends in time series and sales data of similar products. For unstructured data, we use NLP technology to clean text, and then use a custom variant of the pre-trained language model BERT to embed text data into the vector space. For external data, we use NLP technology to extract features and combine them with structured data to obtain the combined structured data. Then, the combined structured data and unstructured data are feature extracted and mapped to a unified feature space to obtain a unified feature vector representation z, which is expressed as follows: z=W s ·x s +W u ·h i Among them, x s Represents a vector in structured data, h i Represents a vector in unstructured data, W s Represents the processed structured data, W u represents unstructured data, W s ∈R p×m and W u ∈R p×d , where R p×m and R p×d is the weight matrix, which projects the structured data and unstructured data into the feature space of the same dimension p, and z∈Rp is the fused feature vector.

3. The method for analyzing Internet big data based on artificial intelligence according to claim 2, characterized in that: The key features of the standardized data are extracted and fused into a unified fusion feature vector, specifically including: For structured data x s , according to its specific numerical value and time series characteristics, a linear transformation with a regular term is used for mapping, which is expressed as follows: Among them, z s Represents the mapping vector of structured data, W s ∈R p×m , R p×m It is a linear mapping matrix that maps m-dimensional structured data to p-dimensional unified space. λ is the regularization coefficient, which is used to control the complexity of the model and prevent overfitting. For unstructured data i , using the transformation matrix W u Projecting it into p-dimensional space, it is expressed as follows: z u =W u ·h i +γ(h i ) Among them, z u Represents the mapping vector of unstructured data, γ(h i ) is a noise suppression term, which is calculated through the attention mechanism or context analysis model and filters irrelevant information in the feedback; W u ∈Rp×d,Rp× d is the projection matrix, which maps the unstructured text embedding into the p-dimensional unified feature space; For external data p , a lag effect control term δ(t) is designed. The lag effect control term δ(t) is related to the policy release time and reflects the gradual impact of the policy on the market, which is expressed as follows: z p+1 =W p ·z p ·δ(t) Among them, z p+1 Represents updated external data, W p ∈R p×q , R p×q is the projection matrix; The mapping vector z of the structured data s , the mapping vector z of unstructured data u and updated external data z p+1 Fusion into a comprehensive feature vector z f , which is expressed as follows: With f =α s With s +α u With u +α p With p Among them, α s , α u , α p is the adaptive weight of each modal feature, satisfying α s +α u +α p =1; According to the comprehensive feature vector z f Design a dynamic update mechanism to ensure that the feature vector can be adaptively updated over time, as shown below: Where β is the time decay coefficient, is the current feature vector, The feature vector generated for the most recent data; At the same time, in the dynamic update process, the regularization term θ is added to control the smoothness of the feature vector, which is expressed as follows: Among them, θ is the regularization coefficient, is the Euclidean distance between feature vectors, which is used to control the amplitude of feature vector update to ensure smooth transition.

4. The method for analyzing Internet big data based on artificial intelligence according to claim 1, characterized in that: In step 2, the demand forecast is based on a dynamic demand forecast model, and the dynamic demand forecast model is constructed as follows: Based on the input multidimensional feature vector z f , use a linear layer plus activation function to perform demand forecasting, the formula is as follows: yd=σ(W d ·With f +b d ) Among them, W d is the demand mapping vector, W d ∈R 1×p , R 1×p is a mapping matrix that maps the p-dimensional feature vector to the demand forecast dimension, b d is the bias term, σ is the activation function, y d It indicates the predicted value of the user's future demand, reflecting the possibility of the user purchasing the product in the future; Add a regularization term Ω(z f ) balances the impact of historical data and new data, expressed as follows: Among them, α represents the regularization coefficient, W d represents the mapping matrix, z f represents the comprehensive feature vector; The interest preference analysis model is constructed as follows: A softmax function is used to map users’ preferences for different product categories into probability distributions, as shown below: y i =softmax(W i ·z f +b i ) Among them, W i represents the interest weight, W i ∈R c×p , R c×p is a weight matrix that transforms the multimodal feature z f Mapped to the interest distribution of the cth product category, b i is the bias term; The sentiment analysis network is constructed as follows: Using a tanh activation function, the output sentiment score range is between [-1, 1]. The specific formula is as follows: y f =tanh(W f ·z f +bf) Among them, W f represents the sentiment weight, W f ∈R 1×p , R 1×p is the weight matrix, b f is the bias term, y f is the sentiment score, which indicates the user's satisfaction with the product; yf>0 indicates that the user's feedback on the product is positive, and yf<0 indicates negative feedback; The comprehensive user profile y profile , which is expressed as follows: Among them, yd represents the user's demand forecast, is the interest preference distribution, It is the result of sentiment analysis.

5. The method for analyzing Internet big data based on artificial intelligence according to claim 4, characterized in that: In the dynamic demand forecasting model, a time-varying weight term λ(t) is also introduced to adjust the weight of demand forecasting according to the change of time, which is expressed as follows: Where λ(t) is the time decay function, represents the updated demand forecast, represents the old demand forecast; In the interest preference analysis model, a time decay regularization term p is also introduced to balance historical interests and future potential interests, which is expressed as follows: in, For the latest interest distribution, is the historical interest distribution; at the same time, the policy influence coefficient η is also introduced into the interest preference analysis model p , which is used to enhance the weighted impact of policy changes on interest preferences, is expressed as follows: Among them, η p reflects the impact of policy changes on interest in different product categories; In the sentiment analysis network, the sentiment amplification factor μ is also introduced to amplify the sentiment fluctuation, which is expressed as follows: Among them, μ is the amplification factor. When the user feedback contains extreme emotions, μ will automatically amplify the emotion score to ensure that the user's emotional fluctuations can be fully captured by the system.

6. The method for analyzing Internet big data based on artificial intelligence according to claim 1, characterized in that: The structure of the dynamic knowledge graph is as follows: The nodes include: user node C, product node P, policy node G, and policy node G; The user node C is based on each user portrait y profile As the characteristic information of the user node, it contains the demand prediction y d 、Interest Preference i and emotional feedback f ; Through user portrait y profile As feature input, denoted as h C =f(y profile ), where f is the feature embedding function; In the product node P, each product node represents a medical device product, denoted as p i , represents the embedding h of product features P =f(p i ); The policy node G is an industry policy and market information node, including the time of policy release, content and impact scope, etc., represented by g i , represents the embedding h of policy features G =f(g i ); The supply chain node S represents the node of the supply chain participant, denoted as s i , represents the embedding h of the supply chain participants S =f(s i ); Relationships include: User-Product Relationship R cp , user-policy relationship R cg and user-supply chain relationship R cs ; The user-product relationship R cp The relationship between users and the products they are interested in or have purchased. The edge weight is determined by the user's interest distribution and purchase history; The user-policy relationship R cg Indicates the impact of policies on user needs; The user-supply chain relationship R cs Represents the interactive relationship between users and supply chain partners, including order delivery history, feedback, and cooperation intensity; A dynamic weight mechanism is introduced to construct the adjacency matrix to embed nodes and relationships, which is expressed as follows: The user-product adjacency matrix A cp (i, j), is represented as follows: A cp (i,j)=α1·sim(y i ,p j )+α2·purchase_history(i,j) Among them, sim(y i , p j ) represents the similarity between user interests and product features, purchase_history(i, j) represents the historical purchase records of users and products, and α1 and α2 are balance parameters; User-policy adjacency matrix A cg (i, j), is represented as follows: A cg (i,j)=β1·impact(g j ,y i )+β2·policy_relevance(i,j) Among them, impact(g j ,y i ) represents the potential impact of policy on user demand, policy_relevance(i, j) represents the correlation between policy and user demand, β1 and β2 are used to adjust the weights of different factors; User-supply chain adjacency matrix A cs (i, j), is represented as follows: A cs (i,j)=γ1·order_frequency(i,j)+γ2·logistics_efficiency(i,j) Among them, order_frequency(i, j) represents the order frequency between users and suppliers, and logistics_efficiency(i, j) represents the supply chain.

7. The method for analyzing Internet big data based on artificial intelligence according to claim 6, characterized in that: The graph neural network model is used to learn the complex relationships between nodes in the dynamic knowledge graph, embed and update the nodes, and capture the structural information in the graph, including: In each layer of the graph neural network, the features of the nodes are propagated through the adjacency matrix and interact with neighboring nodes. The features of each node are updated using the following formula: in, represents the features of node i in the lth layer, N(i) is the set of neighbor nodes of node i, and A ij is the relationship weight between nodes i and j in the adjacency matrix, W(1) is the weight matrix of the lth layer, b(1) is the bias term, and σ is the nonlinear activation function; At the same time, the relationship regularization term is introduced to control the complexity of the relationship between nodes, which is expressed as follows: Among them, A R is the adjacency matrix under different relationship types, α is the regularization coefficient, ||·|| F is the Frobenius norm; The dynamic update mechanism based on time decay is expressed as follows: in, is the feature of node i at the current moment, is the feature update brought by the new data, and β is the time decay coefficient, which controls the weight distribution of historical features and new features.

8. The method for analyzing Internet big data based on artificial intelligence according to claim 1, characterized in that: The embedding representation of the user node is combined with the information of the policy node and the product node to predict the user's future needs and market trends, which is performed as follows: Based on the message passing mechanism of graph neural network, the demand prediction of user nodes can be calculated by the following formula: in, is the prediction result of user demand; W d is the weight matrix that converts user embedding into demand prediction; α g is the policy impact weight, impact(g) is the potential impact of the policy on user demand, and h G is policy embedding; p is the product impact weight, h P Embed for products; The method of recommending the most suitable product and the most suitable supply chain partner to the user based on the prediction results is performed as follows: Embed h through user nodes C and product nodes are embedded in h P The similarity calculation and the historical purchase relationship between the user and the product are used to generate the recommendation score, which is expressed as follows: Among them, y rec is the score of product recommendation; W rec is the weight matrix of product recommendation; is the inner product of user embedding and product embedding, which is used to calculate similarity; γ·purchase_history(C, P) is the influence of historical purchase records, where γ is the historical influence coefficient; Supply chain partner recommendation is calculated by the embedding similarity of user nodes and supply chain nodes, as well as order frequency and logistics efficiency, as shown below: Among them, y sc is the supply chain partner recommendation score; W sc is the weight matrix of supply chain recommendations; is the inner product of the embedding of the user and the supply chain node, which is used to calculate the similarity; δ·order_frequency(C, S) represents the order frequency between the user and the supply chain partner, and δ is the order frequency weight; θ·logistics_efficiency(S) is the impact of logistics efficiency on partner recommendation, where θ is the logistics efficiency weight.

9. The method for analyzing Internet big data based on artificial intelligence according to claim 1, characterized in that: The step 5 specifically includes: Step 5.1: Based on demand forecast and product recommendation results, select appropriate marketing channels for different user groups and allocate marketing resources reasonably; Step 5.2: Based on user portraits and market feedback data, an adaptive user clustering model is constructed to cluster users, and marketing resources and user clustering results are combined to generate personalized marketing content; wherein, the adaptive user clustering model embeds the user vector h C Perform clustering to generate multiple user groups; Step 5.3: Design a real-time feedback mechanism to adjust the current marketing strategy and channel resource allocation by comparing the actual feedback data of personalized marketing content with the expected results. At the same time, design a step control mechanism to ensure that the strategy optimization process can converge and avoid strategy fluctuations caused by excessive adjustments.

10. The Internet big data analysis system based on artificial intelligence is characterized by: The system comprises: The heterogeneous data collection module is used to collect heterogeneous data from different data sources in the medical device industry, standardize them, extract key features from the standardized data, and fuse them into a unified comprehensive feature vector z f ; The image building module is used to build the image according to the comprehensive feature vector z f Generate dynamically updated user profiles, including: Step 2.1: Based on the comprehensive feature vector z f A linear layer plus an activation function is used to predict demand. At the same time, a time-varying weight term λ(t) is added after the linear layer to adjust the weight of demand prediction according to the change of time, ensuring that the weight of old data is the smallest while the weight of new data is the largest. Step 2.2: Based on the comprehensive feature vector z f Establish an interest preference analysis model and design the policy impact coefficient η p Combine and calculate the user's interest distribution in different categories of products; Step 2.3: Based on the comprehensive feature vector z f Design sentiment analysis network based on user feedback, generate sentiment scores, and evaluate user satisfaction with the product; Step 2.4: Integrate demand prediction, interest distribution, and sentiment score to generate a comprehensive user portrait. At the same time, design a dynamic update mechanism based on time decay. Whenever new data enters, the portrait will be updated according to the time weight. The dynamic knowledge graph construction module is used to build a dynamically updated dynamic knowledge graph based on the medical device industry and dynamically updated user portraits, so that market changes are kept up to date. Specifically, it includes: Step 3.1: Collect market information, which includes characteristic data of medical device products P = {p1, p2, ..., p n }, policy and market information G = {g1, g2, ..., g m }, and supply chain partner information S = {s1, s2, ..., s k }, build a dynamic knowledge graph based on market information and comprehensive user portraits; Step 3.2: Use the graph neural network model to learn the complex relationships between nodes in the dynamic knowledge graph, embed and update the nodes, and capture the structural information in the graph; Step 3.3: Design a dynamic update mechanism based on time decay. Whenever new data comes in, the dynamic knowledge graph will automatically update according to the new user behavior and external environment. The dynamic knowledge graph analysis module is used to use the knowledge graph to conduct market analysis, provide product demand forecasts and user recommendations, including: Through the embedded representation of user nodes, combined with the information of policy nodes and product nodes, the user's future needs and market trends are predicted, and the most suitable products and supply chain partners are recommended to the user based on the predicted results. The user's future needs and market trends are predicted, the most suitable products and the most suitable supply chain partners are generated into comprehensive market strategy recommendations; The real-time optimization module is used to provide personalized marketing plans and optimize in real time based on comprehensive market strategy recommendations.

Images