Education platform based on multi-dimensional data fusion and education method thereof

Abstract

The invention provides an education platform based on multi-dimensional data fusion and an education method thereof. The method belongs to the technical field of innovation and entrepreneurship education. Related features are extracted, feature vectors are formed, and dimensionality reduction and fusion are carried out; constructing an integrated learning model; the method comprises the following steps: predicting the future development trend of startup projects through a long-short-term memory network architecture in combination with time sequence analysis, analyzing the incidence relation between startup projects by using a graph neural network, and identifying a potential risk propagation path; and personalized teaching resources and suggestions are generated. According to the method, through LSTM-GNN double-engine driving, the bottleneck problem of dynamic prediction and risk association analysis in a traditional education method is solved, end-to-end optimization from data fusion to decision support is achieved, and technical support with perspectiveness, accuracy and interpretability is provided for entrepreneurial education.

Description

Technical Field

[0001] The present invention provides an education platform based on multi-dimensional data fusion and an education method thereof, belonging to the technical field of innovation and entrepreneurship education. Background Art

[0002] In recent years, the education field has gradually introduced data-driven methods to optimize teaching strategies and improve students' innovation ability. Especially in entrepreneurship education, how to predict the development of entrepreneurial projects and identify potential risks through multi-dimensional data analysis has become a research hotspot. Traditional education methods mostly rely on static evaluation models or single-dimensional student data (such as grades, basic behavior records), making it difficult to dynamically capture the complex evolution laws of entrepreneurial projects and their internal and external correlation risks. Summary of the Invention

[0003] The present invention provides an education platform based on multi-dimensional data fusion and an education method thereof to solve the problems mentioned in the above background art:

[0004] An education method based on multi-dimensional data fusion proposed by the present invention, the method includes:

[0005] S1. Obtain data sources;

[0006] S2. Extract relevant features, form feature vectors and perform dimensionality reduction and fusion;

[0007] S3. Construct an ensemble learning model;

[0008] S4. Through a long short-term memory network architecture, combined with time series analysis, predict the future development trend of entrepreneurial projects, and use a graph neural network to analyze the correlation relationships between entrepreneurial projects to identify potential risk propagation paths;

[0009] S5. Generate personalized teaching resources and suggestions.

[0010] An education platform based on multi-dimensional data fusion proposed by the present invention includes a memory, a processor, and a computer program stored on the memory and executable on the memory. The processor executes the program to implement the education method based on multi-dimensional data fusion as described in any one of the above.

[0011] The beneficial effects of the present invention are as follows: by modeling the time series data of entrepreneurial projects (such as user growth and income fluctuations) through the LSTM architecture, long-term dependencies and nonlinear change patterns can be effectively captured. Compared with traditional time series models (such as ARIMA), the prediction error is reduced by about 20%-30%, providing educators with a quantitative reference for future development and assisting in adjusting teaching strategies in advance; by using GNN to perform topological analysis on the association networks (such as resource dependence, competitive relationships, and cooperative links) between entrepreneurial projects, the diffusion paths of potential risks (such as supply chain disruptions and capital chain breaks) can be explored, key node projects can be identified, and educators can be helped to intervene in time to reduce systemic risks caused by the failure of a single project, and the risk warning accuracy rate can be increased to more than 85%; by using autoencoders to reduce the dimension and fuse features of multi-source heterogeneous data (such as student behavior logs, market data, and social networks), combined with S4's LSTM-GNN joint modeling, the semantic correlation between data can be significantly improved, so that the model can still maintain robustness in sparse data scenarios, and the feature representation efficiency can be improved by 40%. Based on the recommendation system and user portraits, combined with the prediction and risk analysis results, teaching content (such as risk prevention and control courses, market development cases) that adapts to students' shortcomings and entrepreneurial needs can be dynamically generated. Experiments show that the survival rate of student projects using this solution in A / B testing is 25% higher than that of traditional methods. The system regularly updates weights through an integrated learning model, as well as real-time time series prediction and network analysis to form a closed-loop feedback mechanism to ensure that the model is quickly iterated as the market environment changes. The adjustment cycle of educational strategies is shortened from the traditional weeks to hours, significantly improving the timeliness of educational resource allocation. Driven by the LSTM-GNN dual engine, the present invention not only solves the bottleneck problem of dynamic prediction and risk association analysis in traditional educational methods, but also realizes end-to-end optimization from data fusion to decision support, providing entrepreneurship education with a forward-looking, accurate and explainable technical support. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a step diagram of the method described in the present invention. DETAILED DESCRIPTION

[0013] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0014] One embodiment of the present invention, as Figure 1 As shown, an educational method based on multidimensional data fusion, the method comprising:

[0015] S1. Obtain the data sources to be collected from different channels and pre-process them according to the characteristics of each data source;

[0016] S2. Extract relevant features from the raw data based on business understanding and domain knowledge, form feature vectors, and perform dimensionality reduction and fusion on the feature vectors through the autoencoder technology in deep learning;

[0017] S3. Combine multiple machine learning algorithms to build an ensemble learning model, regularly retrain the model according to market changes and the introduction of new data, and adjust the variable weights;

[0018] S4. Through the long short-term memory network architecture, combined with time series analysis, predict the future development trends of startup projects, such as user growth, revenue changes, etc., and use graph neural networks to analyze the correlation relationships between startup projects to identify potential risk propagation paths;

[0019] S5. Based on the user portraits of students, use recommendation system algorithms to generate personalized teaching resources and suggestions, and through the A / B test method, compare the performance of students under different teaching strategies, and provide data-driven decision-making support for educators based on the comparison results.

[0020] The working principle of the above technical solution is: obtain the required data sources from different channels; the data sources include students' learning behavior data, course evaluations, learning achievements, teacher evaluations, social interaction data, etc. Preprocess according to the characteristics of each data source:

[0021] Data cleaning: Remove missing values, outliers, and noisy data to ensure the quality and consistency of the data;

[0022] Data formatting: Convert data in different formats (such as text, images, videos, etc.) into a unified format for subsequent analysis;

[0023] Data normalization: Perform standardization or normalization processing on numerical data to ensure the comparability between different features;

[0024] Based on business understanding and domain knowledge, the system extracts relevant features from the raw data and generates feature vectors; the features include information such as students' learning progress, participation, course difficulty, and teacher feedback. Next, use autoencoder technology to perform dimensionality reduction and fusion on these feature vectors:

[0025] Autoencoder: Perform unsupervised learning through the autoencoder model to map high-dimensional features to a low-dimensional space. The autoencoder trains a neural network so that the input data can be effectively compressed and important information can be retained in the low-dimensional space;

[0026] Feature fusion: By fusing feature vectors from different data sources, form a comprehensive feature representation. This step ensures that the multi-dimensional information of the data can be complementary, improving the prediction ability of subsequent models;

[0027] Combine multiple machine learning algorithms, such as random forest, gradient boosting decision tree (GBDT), support vector machine (SVM), etc., to build an ensemble learning model. The ensemble learning model improves the generalization ability and prediction accuracy of the model by combining multiple weak learners:

[0028] Random forest: Use multiple decision trees for prediction and make decisions through voting mechanism or average value;

[0029] GBDT: Gradually optimize the prediction results by constructing multiple gradient-boosted decision trees;

[0030] SVM: Use support vector machines to find the optimal hyperplane for classification or regression tasks;

[0031] As the market changes and new data are introduced, the system will retrain these models regularly and adjust the variable weights of the models according to the changes in the data to ensure that the models always reflect the latest trends. Through long short-term memory network (LSTM) for time series analysis, predict the future development trends of startup projects, such as user growth, revenue changes, etc.;

[0032] LSTM: LSTM is a recurrent neural network specifically used for time series prediction. It can remember information for a long time period and avoid the vanishing gradient problem in traditional RNNs. Through the LSTM model, the system can identify the patterns and trends in time series.

[0033] Graph neural network (GNN): Use GNN to model and analyze the association relationships between startup projects. GNN can effectively capture the dependencies between nodes and identify potential risk propagation paths. For example, the failure of a certain startup project may affect the success of other related projects.

[0034] Based on the user profiles of students, the system uses recommendation algorithms to generate personalized teaching resources and suggestions. For example, recommend suitable courses, learning materials or practice questions to help students achieve better results in learning:

[0035] User profile: Build a personalized profile of students based on information such as their learning behaviors, grades, interests, etc.

[0036] Recommendation system: Generate personalized recommendation content through methods such as collaborative filtering, matrix factorization or deep learning.

[0037] In addition, through the A / B test method, compare and analyze the performance of students under different teaching strategies to obtain data support on the effectiveness of teaching strategies:

[0038] A / B Testing: The system divides students into two groups, uses different teaching strategies respectively, and compares the differences in learning performance between the two groups of students. Based on this data, educators can make more scientific and data-driven decisions. Through these methods, educators can optimize teaching strategies according to the data analysis results and improve students' learning effects.

[0039] The effects of the above technical solutions are as follows: By collecting data from different channels and combining the autoencoder technology of deep learning for feature extraction, dimensionality reduction and fusion, the system can comprehensively and accurately capture the correlations of various factors in the education process, improve the effectiveness and comprehensiveness of the data, and help achieve more accurate analysis and decision-making; Based on the user portraits of students and the recommendation system algorithm, personalized learning resources and suggestions can be generated for each student, optimizing the learning path of students and improving learning effects. By analyzing students' behavior data, the system can accurately recommend the most suitable learning content, thereby improving students' learning interest and motivation; By combining multiple machine learning algorithms such as random forest, GBDT and SVM, an ensemble learning model is constructed, effectively improving the prediction accuracy and robustness of the system. Regularly retraining and adjusting the model helps ensure that the system can adapt to market changes and the introduction of new data, ensuring that the prediction results always match the actual situation; With the help of long short-term memory network (LSTM) and graph neural network (GNN), accurate predictions can be made for factors such as students' learning achievements and project development trends in the education system. At the same time, the correlation relationships between students, courses and other factors can be analyzed through the graph neural network to identify potential risk propagation paths, providing forward-looking guidance for education decision-making; The A / B testing method provides real data support for the effects of different teaching strategies, helping educators compare and analyze the impacts of different strategies on students' performance, thereby making data-driven decisions, optimizing teaching content and methods, and improving teaching effects; Through personalized recommendation and accurate prediction, the system can help educational institutions make efficient use of resources, accurately match students' needs with resource supply, avoid resource waste, and improve the overall efficiency of education; By applying advanced technologies such as deep learning and machine learning, the system can respond to changes in the education environment in real time, quickly adapt to the diverse needs of students, and provide intelligent and dynamically adjustable solutions for educators and managers.

[0040] In one embodiment of the present invention, S1 includes:

[0041] S11. Determine the data types to be collected, and select the corresponding data source channels according to the data types;

[0042] S12. Preprocess the collected raw data, uniformly convert data in different formats into a format suitable for subsequent processing, and use natural language processing technology (NLP) to process social media data to extract key information;

[0043] S13. Preprocess the financial statement formats of different enterprises using a standardized method, integrate the preprocessed data into a unified data warehouse, and perform data verification and consistency checks.

[0044] The working principle of the above technical solution is as follows: Analyze the types of data required by the system, which may include user behavior data, enterprise financial data, social media content, market trends, etc. Determine the types and characteristics of the required data according to the goal. For example, if the goal is to analyze user sentiment and market feedback, social media data and comments need to be collected; if the goal is to conduct enterprise financial analysis, financial statements need to be collected; Select appropriate data sources according to the requirements of different data types. For example, social media data can be obtained through API interfaces; financial statement data can be obtained through enterprise annual reports, financial report databases, or public financial data interfaces; For each data source, select an appropriate collection method. For example, social media data may need to be crawled through a crawler program or obtained regularly through an API interface; financial statement data may be crawled or imported regularly by establishing connections with various big data providers; For the collected raw data, it is first necessary to uniformly convert it into a standard format suitable for subsequent processing (such as CSV, JSON, XML, etc.) to ensure that the data is convenient for subsequent operations in a unified format. For example, social media comment data may be unstructured and need to be stored in JSON format after being crawled by a crawler; while financial statements may be tabular PDF files and need to be converted into CSV format; For social media data (such as text content like user comments, posts, etc.), use NLP technology for processing:

[0045] Text cleaning: Remove irrelevant information, special characters, or noisy data to ensure the accuracy of subsequent analysis.

[0046] Word segmentation: Split sentences into individual words for word segmentation so that the machine can understand the basic structure of the text.

[0047] Part-of-speech tagging: Determine the grammatical role of each word (such as noun, verb, adjective, etc.) through part-of-speech tagging to help identify the semantics of the sentence.

[0048] Named entity recognition (NER): Identify entity information (such as company names, person names, locations, etc.) in the text to help extract key events or topics.

[0049] Sentiment analysis and topic tagging: Judge the sentiment tendency (such as positive, negative, or neutral) in user comments or posts through sentiment analysis. At the same time, extract popular topic tags related to the comment content to provide support for data classification and analysis.

[0050] The financial statements of different enterprises may adopt different formats, and the data may be presented in different column structures or units (such as ten thousand yuan, thousand yuan). The goal of standardization processing is to unify these data formats for subsequent analysis. For example, key financial indicators such as revenue, profit, and cash flow may use different terms and structures in the statements of different enterprises. Standardization processing can unify these terms and standardize them; before data integration, it is necessary to perform consistency verification on the data from various sources. Check whether the data is complete, whether there are missing or abnormal values, and ensure that the data between different enterprises can be unified and the data quality is ensured. In this step, data deduplication, data format checking, etc. may also be performed to ensure that the finally integrated data can be used as high-quality input; Integrate the processed data (including social media text data and enterprise financial data) into a unified data warehouse. A data warehouse is a centralized database that provides data support for subsequent analysis and decision-making. During the integration process, data cleaning, deduplication, noise removal, etc. are usually performed to ensure high-quality and consistent final data.

[0051] The effects of the above technical solutions are as follows: By clarifying the data type and selecting appropriate data source channels, it is possible to ensure that the collected data has high relevance and quality, thus providing strong support for subsequent data analysis. The multi-channel data collection method can cover a wider range of information sources, enhancing the representativeness and comprehensiveness of the data; The data preprocessing step unifies and converts the data format, enabling data from different sources and forms to be processed under a unified standard, avoiding analysis deviations or errors caused by inconsistent data formats, and improving the compatibility and operability of the data; Using natural language processing technology (NLP) to clean, segment, part-of-speech tag, named entity recognition, etc. of social media text data can extract valuable key information, such as user comments, sentiment tendencies, topic tags, etc. This information can provide rich input data for applications such as sentiment analysis, market trend prediction, and brand monitoring.; By standardizing the format of financial statements, it is ensured that financial data from different enterprises can be compared and analyzed on the same platform. Standardization processing makes financial indicators (such as revenue, profit, cash flow, etc.) consistent, avoiding incorrect analysis or misleading caused by inconsistent formats; Integrating the preprocessed data into a unified data warehouse facilitates centralized management and efficient storage. The steps of data verification and consistency checking ensure data quality, reduce analysis deviations caused by data errors or omissions, and improve the accuracy and reliability of analysis results; By integrating and standardizing different types of data (such as social media comments and financial data), this technical solution provides rich and accurate input data support for subsequent data analysis and decision-making. It can help enterprises conduct comprehensive analysis from multiple dimensions such as user sentiment, market trends, and financial conditions, providing a scientific basis for strategic decision-making.

[0052] In one embodiment of the present invention, S2 includes:

[0053] S21. Extract relevant features from the original data based on business understanding and domain knowledge, and filter the extracted features through a feature selection algorithm to remove redundant features;

[0054] S22. Construct a feature vector. For numerical features, directly convert them into feature vectors. For non-numerical features, use an encoding method for conversion to form numerical feature vectors; and perform normalization processing on the feature vectors;

[0055] S23. Utilize the autoencoder technology in deep learning to construct a dimensionality reduction and fusion model for the feature vector, use the preprocessed feature vector to train the autoencoder model, and optimize the model performance by adjusting the model parameters;

[0056] S24. Through the trained autoencoder model, perform dimensionality reduction and fusion on the feature vector to form a low-dimensional representation vector.

[0057] The working principle of the above technical solution is as follows: based on the understanding of specific businesses and industries, features related to problem solving are extracted from the original data. These features may include market demand scale, technology maturity, team experience, capital adequacy ratio and industry trends, etc. These features play an important role in prediction and analysis; in order to improve the effect of the model and avoid redundancy, feature selection algorithms (such as chi-square test, mutual information, recursive feature elimination, etc.) are used to screen the extracted features and remove irrelevant or redundant features. This step helps to reduce the data dimension, improve the efficiency of model training, and reduce the risk of overfitting; for numerical features, they are directly converted into feature vectors. These features can directly reflect information and can be used with other numerical features; for non-numerical features such as text descriptions and category labels, corresponding encoding methods (such as unique hot encoding, word embedding, label embedding, etc.) are used to convert them into numerical feature vectors. This ensures that all features are input into subsequent models in a unified format; all feature vectors are normalized to scale feature values ​​in different ranges to a unified standard range (such as [0,1] or [-1,1]). This helps avoid deviations during model training caused by large differences in eigenvalues. Autoencoder is an unsupervised learning method that can automatically learn an effective representation of input data. In this step, the autoencoder is used to reduce the dimension and fuse the feature vector, mapping the high-dimensional features to a low-dimensional latent space. The autoencoder consists of two parts: an encoder and a decoder. The encoder maps the input data to a low-dimensional space, and the decoder reconstructs the low-dimensional representation into the original data. The autoencoder is trained using the preprocessed feature vectors, and the model performance is optimized by adjusting the model parameters (such as the number of layers, the number of nodes, the learning rate, etc.). The optimization process is usually performed using gradient descent or its variants (such as the Adam optimizer) to minimize the reconstruction error or other loss functions. The high-dimensional feature vector is mapped to a low-dimensional latent space through the trained autoencoder model. This low-dimensional representation vector retains the key information of the original data and retains the main features of the data as much as possible while reducing the dimension. The feature vector after dimensionality reduction is more suitable for subsequent machine learning or deep learning tasks, which not only improves computational efficiency but also reduces the complexity of the model.

[0058] The effects of the above technical solutions are as follows: By extracting relevant features based on business understanding and domain knowledge, and combining with a feature selection algorithm to remove redundant features, the prediction accuracy of the model can be significantly improved. The optimized feature set can better reflect the essence of the target problem, which helps the model learn quickly and improve the prediction accuracy; Using an autoencoder for dimensionality reduction and fusion of feature vectors can effectively reduce the data dimension while retaining key information. This not only reduces the complexity of the model but also significantly improves the computational efficiency, making the training process more efficient and reducing the demand for hardware resources; This solution can handle various data types such as numerical, categorical, and text. Through feature vector construction and encoding methods (such as one-hot encoding, word embedding, etc.), various types of input data can be converted into a unified format to ensure that the model can handle complex heterogeneous data; The feature selection and dimensionality reduction processes help reduce the risk of overfitting, especially in the case of high data dimensions. By removing redundant features and effectively reducing the dimension of features, the model can better adapt to unseen data and improve the generalization ability of the model; As an unsupervised learning method, the autoencoder has strong flexibility. By adjusting the parameters of the autoencoder (such as the number of layers, the number of nodes, and the learning rate, etc.), the dimensionality reduction effect can be optimized, and further improve the model performance, making this solution adaptable to different data sets and problem requirements; The low-dimensional feature vectors after dimensionality reduction retain the core information of the original data, enabling the model to be trained on a more concise and interpretable representation. This low-dimensional representation not only helps improve the execution efficiency of the model but also makes subsequent analysis and decision-making more transparent and easy to understand.

[0059] In one embodiment of the present invention, the S23 includes:

[0060] Design an autoencoder with multiple hidden layers, and each layer undertakes different feature extraction and dimensionality reduction tasks. By stacking multiple encoding and decoding layers, high-level feature representations are gradually abstracted;

[0061] Introduce a sparsity constraint in the encoding layer;

[0062] Introduce cross-layer connections between the hidden layers of the autoencoder to allow direct interaction of feature information at different levels; introduce an attention mechanism in the decoding layer to dynamically adjust the contribution of different features during the reconstruction process;

[0063] Adopt grid search to finely tune the hyperparameters of the autoencoder, and use the Xavier or He initialization method to initialize the weights of the autoencoder;

[0064] Introduce an L2 regularization term in the loss function. At the same time, by adding a Dropout layer, randomly discard some neurons during the training process; and monitor the loss or accuracy metrics on the validation set during the training process. When the metrics no longer improve, stop training in advance;

[0065] Add noise to the input data, and then train the autoencoder to recover the original features from the noisy data. Evaluate the dimensionality reduction effect of the autoencoder model by calculating the reconstruction error between the feature vectors after dimensionality reduction and the original feature vectors, as well as the performance improvement of the features after dimensionality reduction in tasks such as classification and regression.

[0066] According to the evaluation results, adjust the architecture, parameters, regularization strategy, etc. of the autoencoder to form a closed-loop optimization process.

[0067] The working principle of the above technical solution is as follows: An autoencoder is an unsupervised learning algorithm, usually consisting of two parts: an encoder and a decoder. The encoder maps the input data into a low-dimensional latent space (i.e., feature vectors), while the decoder attempts to reconstruct the original input data from these low-dimensional feature vectors; This solution designs an autoencoder with multiple hidden layers, and each layer is responsible for different levels of feature extraction and dimensionality reduction tasks. By stacking multiple encoding and decoding layers, the model can gradually abstract more high-level and complex feature representations; Cross-layer connections are introduced between the encoding layers (similar to the skip connections in ResNet), allowing feature information to directly interact between different levels, thus avoiding the vanishing gradient problem and enhancing the model's feature learning ability; A sparsity constraint (such as L1 regularization or KL divergence) is introduced in the hidden layer of the encoder, aiming to encourage the model to make the activation values of some neurons zero as much as possible when learning features, so as to obtain sparse feature representations. This helps to reduce redundant information and ensure that the model learns the most discriminative features; An attention mechanism is introduced in the decoding layer, enabling the model to dynamically adjust the contribution of different features to the final output during the reconstruction process. The attention mechanism can help the autoencoder focus more effectively on important features and improve the reconstruction accuracy; Grid search is used to finely tune the hyperparameters of the model (such as the number of layers, number of nodes, learning rate, batch size, etc. of the autoencoder). By traversing different hyperparameter combinations, the best model configuration is found to ensure that the model can achieve optimal performance in a given task; Initialization method: Use the Xavier or He initialization method to initialize the weights of the network, which helps to alleviate the problems of vanishing gradients or exploding gradients, thereby improving the stability of training; An L2 regularization term is added to the loss function to control the complexity of the model and reduce the risk of overfitting. In addition, a Dropout layer is added during the training process to further reduce overfitting by randomly discarding some neurons, ensuring that the model can have better generalization ability on unknown data; To improve the robustness of the model, noise is added to the input data, and then the autoencoder is trained to recover the original features from the noisy data. By calculating the reconstruction error (i.e., the difference between the dimensionality-reduced feature vectors and the original feature vectors), the performance of the model in the dimensionality reduction task is evaluated. In addition, by monitoring the performance improvement of the autoencoder in specific tasks (such as classification, regression), its dimensionality reduction effect is further evaluated; According to the evaluation results of the model performance, the autoencoder architecture, parameter settings, regularization strategies, etc. are continuously adjusted to form a closed-loop optimization process. Each adjustment will evaluate the effect of the model through the training and validation processes and make corresponding modifications according to the feedback, finally forming an optimal autoencoder model; During the training process, continuously monitor the loss or accuracy metrics on the validation set. When these metrics no longer improve, trigger the early stopping mechanism to avoid overtraining and reduce the training time. This helps to save computing resources and prevent overfitting.

[0068] The effects of the above technical solutions are as follows: By stacking multiple encoding and decoding layers, higher-level feature representations are gradually abstracted. This enables the model to extract richer and more accurate feature information from the original data, providing more meaningful inputs for subsequent tasks (such as classification and regression); Introducing cross-layer connections (such as skip connections in ResNet) allows feature information to flow directly between different levels, solving the problems of vanishing gradients or information loss, enhancing the feature expression ability, and improving the training efficiency of the model; Sparsity constraints (such as L1 regularization or KL divergence) make the activation of some neurons zero by restricting the model during feature learning, promoting the model to learn more sparse feature representations. This can not only improve the interpretability of the model but also help reduce redundant information and improve the generalization ability of the model; By introducing the attention mechanism, the model can dynamically adjust the importance of different features during the reconstruction process, thereby improving the accuracy of the autoencoder in the feature reconstruction task. This means that the model can better recover the original features from noisy data and give higher weights to key features; By carefully adjusting hyperparameters such as the number of layers, the number of nodes, and the learning rate through grid search, the model configuration most suitable for the current task can be found, thereby improving the model performance, reducing the risk of overfitting, and optimizing the training time; Using the Xavier or He initialization method can effectively alleviate the problems of vanishing or exploding gradients and improve the training stability; Adding an L2 regularization term to the loss function and using a Dropout layer during training can effectively reduce overfitting and improve the generalization ability of the model on unknown data. By avoiding the model from relying too much on certain specific features in the training data, the robustness of the model is further improved; By adding noise to the input data and training the model to recover the original features from the noise, the robustness of the autoencoder can be enhanced, making it perform more robustly when dealing with incomplete or noisy data in practical applications; According to the model evaluation results, repeated adjustments are made to the architecture, parameters, regularization strategies, etc., forming a closed-loop optimization process. This continuous optimization mechanism enables the model to continuously improve during the training process and finally achieve higher performance; By calculating the reconstruction error and combining the performance improvement of tasks such as classification and regression, the dimensionality reduction effect of the autoencoder and its effectiveness in practical applications can be clearly evaluated. A lower reconstruction error and better downstream task performance indicate that the autoencoder can provide high-quality feature representations.

[0069] In one embodiment of the present invention, S3 includes:

[0070] S31. Combine multiple machine learning algorithms as the base learners of the ensemble learning model; adopt an ensemble strategy to combine multiple base learners into an ensemble learning model; and perform parameter tuning on each base learner in the ensemble learning model;

[0071] S32. Divide the preprocessed data into a training set, a validation set, and a test set, and use the training set to train the ensemble learning model;

[0072] S33. Use the validation set to validate the trained model, evaluate the performance of the model, and iteratively optimize the model according to the validation results;

[0073] S34. Use the feature importance evaluation method to determine the contribution degree of each feature to the prediction result of the ensemble learning model, and adjust the variable weights in the ensemble learning model according to the feature importance evaluation results. The contribution degree is obtained through the following formula:

[0074]

[0075] Where, W f represents the weight of feature f; I f·i represents the importance score of feature f in the i-th base learner; N represents the number of base learners; C f represents the number of base learners that consider feature f important; represents the standard deviation of the importance scores of feature f; represents the average value of the importance scores of feature f; ∈ is a very small positive number used to prevent division by zero errors.

[0076] The working principle of the above technical solution is as follows: Select different types of base learners (such as random forests, gradient boosting decision trees (GBDT), support vector machines (SVM), etc.). These base learners each have different characteristics and advantages and can model the dataset in different ways. By selecting multiple base learners, the diversity of the model can be increased, the bias can be reduced, and the generalization ability can be improved; By resampling the training data, training multiple base learners, and summarizing their prediction results (usually using voting or averaging methods). Bagging can reduce the variance of the model and avoid overfitting; Sequentially train multiple base learners in a weighted manner, and each new base learner will focus on the samples that the previous base learner did not correctly predict. The Boosting strategy helps to reduce the bias of the model and increase the prediction accuracy of the model; For each base learner, parameter tuning is required to ensure that each model works in its optimal configuration. This can be done through cross-validation, grid search, etc. The tuned base learner can perform more precisely and efficiently on the training set; Before this step, preprocess the data, such as filling in missing values, normalizing, standardizing, etc. The quality of the data directly affects the effect of the model, so the preprocessing step is crucial; Divide the dataset into three parts:

[0077] Training set: Used to train the ensemble learning model, and the model learns the rules of the data from it.

[0078] Validation set: Used to evaluate the performance of the model during training and perform hyperparameter tuning.

[0079] Test set: Used to evaluate the generalization ability and actual effect of the final model to ensure that the model is not overfitting.

[0080] Train an ensemble learning model using the training set. During training, multiple base learners will use the dataset to build models and form an ensemble model. The training process includes calculating the loss function, optimizing model parameters, etc.; Evaluate the performance of the trained model using the validation set. Common evaluation metrics include accuracy, precision, recall, F1-score, etc. If the model performs poorly, return to adjust the model's parameters or optimize feature selection; According to the performance evaluation results on the validation set, further adjust the model. For example, the learning rate, the number of base learners, the ensemble strategy (Bagging or Boosting), etc. can be adjusted; The goal of iterative optimization is to ensure that the final model has good generalization performance and is not prone to overfitting or underfitting through repeated training and adjustment; After training is completed, evaluate the importance of each feature to the model's prediction results; including:

[0081] Gini coefficient: Used to measure the contribution of each feature in a decision tree to the data classification result.

[0082] Information gain: Calculate the information gain of each feature after data splitting to help judge the importance of the feature.

[0083] SHAP value: SHAP (Shapley Additive Explanations) is to evaluate the contribution of each feature to the final model output through game theory methods, and can provide local explanations for each sample.

[0084] Based on the evaluation results of feature importance, the weights of each feature in the model can be adjusted. For example, according to the importance ranking, select to remove low-importance features, or increase the weights of high-importance features. This step can further improve the performance of the model, reduce unnecessary calculations and reduce the risk of overfitting.

[0085] The effects of the above technical solutions are as follows: By combining multiple machine learning algorithms (such as random forest, GBDT, SVM, etc.), each base learner can understand the data from different perspectives, reducing the bias and overfitting risks that may occur in a single model. Ensemble learning helps improve the accuracy and robustness of the overall model by integrating the prediction results of multiple base learners; adopting ensemble strategies (such as Bagging and Boosting) can effectively reduce the variance and bias of the model, thereby improving the generalization ability of the model. Bagging avoids overfitting by reducing variance; Boosting reduces bias and improves prediction performance by gradually improving the model; tuning the parameters of the base learners enables each base learner to work under the optimal configuration, thus enhancing the performance of the overall ensemble learning model. Automatically adjusting the model's parameters to make it more adaptable to different data characteristics and problem requirements, improving the adaptability of the model; by dividing the data into training set, validation set, and test set, the performance of the model can be better evaluated, avoiding overfitting, and ensuring the reliability and effectiveness of the final model in practical applications. The use of the validation set helps adjust the model and avoid over-reliance on the training set; through multiple validations and iterative optimizations of the model, the prediction accuracy can be continuously improved, possible biases can be corrected, and the stability of the ensemble learning model under different environments and conditions can be ensured; using feature importance evaluation methods (such as Gini coefficient, information gain, SHAP value, etc.) to identify and adjust key features, making the model more focused on the most important features, reducing unnecessary computational volume and dimensions, thereby enhancing the efficiency and performance of the model; feature importance evaluation (such as SHAP value) can help understand the specific contributions of each feature to the model output, contributing to providing the transparency of model decision-making, making the model more interpretable, and facilitating subsequent optimization and improvement; in the process of feature selection and weight adjustment, removing low-importance features can reduce unnecessary calculations, reduce the complexity of the model, and thus reduce the consumption of computing resources. The above contribution degree calculation formula not only considers the average importance score of each feature in all base learners, but also introduces the ratio of the standard deviation to the average value of the feature importance scores, which helps identify those features that are stable and important in different base learners; by introducing the ratio of the standard deviation to the average value, the formula can distinguish those features with large fluctuations in importance scores in different base learners, thereby improving the stability of feature selection; introducing a very small positive number ∈ effectively prevents division-by-zero errors and enhances the robustness of the formula; The item reflects the proportion of base learners that consider feature f important, which helps identify features commonly considered important by most base learners, thus enhancing the consensus of feature selection; through the calculated feature weights, one can intuitively understand the contribution degree of each feature to the prediction result of the ensemble learning model, providing a strong basis for subsequent adjustment of variable weights; the weight adjustment based on feature importance helps optimize the model structure, enabling the model to pay more attention to features that have an important impact on the prediction result, thereby improving the prediction performance and generalization ability of the model.

[0086] In one embodiment of the present invention, the S34 includes:

[0087] Evaluate the importance of features from multiple dimensions (such as statistical significance, model dependence, marginal contribution, etc.) through multiple evaluation methods (such as Gini coefficient, information gain, SHAP value, etc.); the feature importance is evaluated by the following formula:

[0088]

[0089] where FIS represents the feature importance score; N, M, and L respectively represent the number of individual feature importance evaluation methods, feature interaction effect evaluation methods, and feature clustering impact evaluation methods; F i represents the score of the i-th individual feature importance evaluation method, such as Gini coefficient, information gain, etc.; w i represents the weight of the i-th individual feature importance evaluation method; α i is the exponential weight representing the i-th individual feature importance evaluation method, used to adjust the influence of this evaluation method; I j represents the score of the j-th feature interaction effect evaluation method, such as interaction effect analysis based on SHAP value; w' j represents the weight of the j-th feature interaction effect evaluation method; γ j represents the exponential weight of the j-th feature interaction effect evaluation method; C k represents the score of the k-th feature clustering impact evaluation method, such as the centrality or influence of the feature in the clustering; w″ k represents the weight of the k-th feature clustering impact evaluation method; γ k represents the exponential weight of the k-th feature clustering impact evaluation method; R t represents the score of feature redundancy over time t, which can be a measure of the correlation or similarity between features; λ represents the weight of feature redundancy, usually negative, used to penalize redundant features; max(F i ), max(I j ), max(C k ), max(R t) represent the maximum possible values of the scores of each evaluation method respectively; ∈ represents a very small positive number used to prevent division by zero error; min(F i ) and min(I j ) and min(C k ) and min(R t ) represent the minimum values of the scores of each evaluation method respectively;

[0090] Analyze the interaction effects between features through SHAP values to identify which feature combinations can significantly improve the prediction ability of the model;

[0091] Group features using a clustering algorithm based on the correlation or importance scores between features;

[0092] Based on the feature grouping, use a hierarchical model to dynamically adjust the weights of features within and between groups;

[0093] Introduce an online learning or incremental learning mechanism, and enable the model to dynamically adjust feature weights as new data arrives through a time-window-based weight update strategy;

[0094] Regularly evaluate the current importance of all features, eliminate low-contribution features according to thresholds or the degree of performance impact, and quickly evaluate newly introduced features at the same time;

[0095] Generate a force-directed graph using SHAP values to visually display the positive or negative impacts of features on the prediction results and the interactions between features;

[0096] Analyze the contribution distribution of each feature under different prediction results to identify the behavior patterns of features in different situations;

[0097] Through a model performance monitoring system, regularly evaluate the performance of the model on new data, and according to the monitoring results, timely adjust feature weights or introduce new features to form a closed-loop optimization process.

[0098] The working principle of the above technical solution is as follows: Methods such as the Gini coefficient, information gain, and SHAP values are used to evaluate the contribution of features to the model from multiple dimensions; The Gini coefficient (Gini Impurity) evaluates the "purity" contribution of features in classification problems. The lower the Gini coefficient of a feature, the greater the contribution of the feature to the classification of the decision tree model; Information gain measures the effectiveness of feature partitioning of data. The greater the information gain, the greater the contribution of the feature in model training, and it is commonly used in models such as decision trees.

[0099] SHAP values (Shapley Additive Explanations), based on the idea of game theory, can explain the contribution degree of each feature to the model output. Especially in the application of complex models (such as random forests, XGBoost, etc.), it makes the decision-making process of the model more interpretable. These methods reveal which features have a significant impact on the model prediction results by calculating the importance scores of features, thus helping to further adjust and optimize the model. Analyze the interaction effects between features using SHAP values to identify which combinations of features can significantly improve the model's prediction ability. SHAP values can calculate the contribution of a single feature and the joint contribution of multiple feature combinations to the model prediction results. This kind of analysis helps to reveal which features can improve the model performance when combined together, or which features are redundant and can be merged or deleted. Group features using clustering algorithms (such as K-means, hierarchical clustering) based on feature correlation or importance scores. Clustering algorithms group features with strong correlation together by calculating the similarity between features, which can reduce redundancy between features and improve the training efficiency of the model. The clustering results can provide a basis for subsequent hierarchical models, enabling features within the same group to be jointly optimized according to their mutual relationships. On the basis of feature grouping, use hierarchical models (such as the hierarchical structure of decision trees, random forests) to dynamically adjust the weights of features within and between groups. Hierarchical models adjust the role of features more meticulously by constructing a multi-level decision-making structure and combining the contributions of features within each feature group. This method can dynamically adjust the weights according to the performance of different feature groups at each level, enhancing the model's adaptability to complex data relationships. Introduce an online learning or incremental learning mechanism to enable the model to dynamically adjust feature weights as new data arrives. Online learning can continuously update the model parameters and feature weights according to the latest data when dealing with streaming data. The incremental learning mechanism avoids retraining the entire model by incrementally adding new data during the model training process, improving the model update efficiency. Eliminate features with little impact on the model prediction results by setting thresholds to reduce unnecessary computational overhead. Quickly evaluate newly introduced features to determine their impact on model performance, so as to decide whether to retain or adjust them in a timely manner. Generate a force-directed graph using SHAP values to visually display the positive or negative impacts of features on the prediction results and the interactions between features. The force-directed graph can display the relationships and interaction effects between features, helping users understand how each feature affects the final prediction result, especially having significant advantages in the analysis of complex models. Through graphical means, non-technical personnel can also understand the model's decision-making process, improving interpretability.Analyze the contribution distribution of each feature under different prediction results (such as classification labels, regression value intervals) to identify the behavioral patterns of features in different situations; different features may have different performances in different prediction situations. By analyzing these situations, it is possible to identify which features have stronger prediction capabilities in certain cases and further optimize the model; through the model performance monitoring system, regularly evaluate the performance of the model on new data. According to the monitoring results, timely adjust the feature weights or introduce new features to form a closed-loop optimization process; the monitoring system tracks the prediction effect of the model in real time, analyzes the performance of features, and adjusts the model in a timely manner according to changes; through the closed-loop optimization mechanism, ensure that the model continuously and efficiently adapts to new data distributions and task requirements.

[0100] The effects of the above technical solutions are as follows: Through various feature importance evaluation methods (such as Gini coefficient, information gain, SHAP value, etc.), features can be analyzed in detail from multiple dimensions, and the features that contribute most to the prediction can be identified. This helps to eliminate redundant or irrelevant features, improve the accuracy and stability of the model; SHAP value analysis further reveals the interaction between features, ensuring the optimization of feature combinations. Feature combinations can significantly improve the prediction ability of the model, reduce errors, and enhance the adaptability to complex data patterns; Based on the importance scores and correlations of features, clustering algorithms (such as K-means, hierarchical clustering) are used to group features, thereby reducing redundancy and improving training efficiency. Feature clustering ensures more consistent processing of similar features, helping to improve the overall performance of features; On the basis of grouping, a hierarchical model (such as decision tree, random forest, etc.) is used for dynamic adjustment, so that the influence of different feature groups can be accurately controlled. The hierarchical model can better capture the non-linear relationship between complex features, improving the flexibility of the model; By introducing an online learning or incremental learning mechanism, the model can be adjusted in real time according to new data, avoiding the disadvantage of having to retrain the model in traditional batch learning methods. Dynamically adjusting feature weights can effectively cope with changes in the data stream, enabling the model to maintain good prediction ability; The time window based on the weight update strategy helps the model to continuously optimize in a changing environment, further enhancing the adaptive ability of the model; By regularly evaluating the current importance of features and eliminating low-contributing features in combination with thresholds or performance impact levels, the computational complexity and memory consumption can be reduced, improving the running efficiency of the model; The newly introduced features are quickly evaluated, and it can be judged in time whether they have a positive impact on the performance of the model, so as to make adjustments or optimizations. Effective feature screening can improve the generalization ability of the model and avoid overfitting; The force-directed graph generated by using the SHAP value intuitively shows the positive and negative effects of each feature on the prediction result, and reveals the interaction between features. This graphical display can help data scientists or non-technical personnel understand the decision-making process of the model, improving the transparency and trust of the model; By analyzing the behavior patterns of features under different prediction results, the contributions and performances of features in different situations can be identified, further enhancing the interpretability of the model; The model performance monitoring system can regularly evaluate the performance of the model on new data and timely detect the performance decline or optimization space of the model. Through real-time adjustment of the weights of new features and existing features, a closed-loop optimization mechanism is formed to ensure that the model is always in the best state; The closed-loop optimization process can quickly respond and adjust when the data distribution or task requirements change, ensuring long-term stable prediction effects; Feature screening, feature clustering, and dynamic adjustment mechanisms significantly reduce unnecessary calculations and improve the utilization efficiency of computing resources. By streamlining the feature set and dynamically optimizing feature weights, the time and memory consumption of model training are reduced, thus improving the overall efficiency.By integrating different types of feature importance evaluation methods (such as Gini coefficient, information gain, SHAP value, etc.), the above formula can comprehensively evaluate the importance of features from multiple dimensions such as statistical significance, model dependence, and marginal contribution; it takes into account feature interaction effects (such as interaction effect analysis based on SHAP values) and feature clustering effects (such as the centrality or influence of features in clusters), thus providing a deeper understanding of the role of features in the model; the weights and exponential weights in the formula allow users to adjust the contribution degrees of different evaluation methods according to specific application scenarios and requirements; by adjusting these weights and exponential weights, the relative importance between different evaluation methods can be flexibly balanced to adapt to different datasets and models; a very small positive number is introduced to prevent division-by-zero errors, enhancing the numerical stability of the formula; by normalizing the scores to the range of [0, 1] (by subtracting the minimum value and dividing by the difference between the maximum value and the minimum value plus ∈), the problem of inconsistent score dimensions of different evaluation methods is reduced, improving the comparability of evaluation results; the feature redundancy term and its weight λ in the formula allow for penalizing redundant features, which helps reduce noise and unnecessary complexity in the model; by reducing the scores of redundant features, the generalization ability and interpretability of the model can be improved; by combining multiple evaluation methods and dimensions, the formula provides richer feature importance information, helping to explain the basis for model decisions; by analyzing the scores and weights of different evaluation methods, it is possible to deeply understand which features are most critical for model prediction and how they interact with each other.

[0101] In one embodiment of the present invention, step S4 includes:

[0102] S41. Input the preprocessed data into the trained ensemble learning model, and use the long short-term memory network architecture in combination with time series analysis to predict the future development trend of the entrepreneurial project; analyze the prediction results to identify potential growth points and challenges;

[0103] S42. Use a graph neural network to analyze the association relationships between entrepreneurial projects to identify potential connections and influences between projects;

[0104] S43. Based on the results of the project association analysis, identify potential risk propagation paths, and provide early warning information for entrepreneurs according to the risk propagation paths.

[0105] The working principle of the above technical solution is as follows: The preprocessed data contains time series data of key indicators such as historical user growth, revenue changes, and market share; these data go through steps such as cleaning and standardization to remove noise and fill in missing values to ensure data quality; the preprocessed time series data is input into an ensemble learning model. LSTM (Long Short-Term Memory network) is a neural network structure specifically designed for processing sequence data, capable of effectively capturing long-term dependencies in time series; LSTM controls the flow of information through three gates (input gate, forget gate, output gate), which can retain key information in the time series while filtering out irrelevant information; after being trained, the LSTM model can predict future development trends based on historical data, specifically including indicators such as user growth, revenue changes, and market share. Based on the prediction results of LSTM, trend analysis is carried out to identify potential future growth points and possible challenges. For example, if the market share growth of a startup project slows down, it may be a potential future challenge; the prediction results help entrepreneurs make more accurate decisions and formulate development strategies. According to the business nature, market domain, investor relations, partners, etc. of the startup project, a connection graph between startup projects is constructed. Each project serves as a node in the graph, and the relationships between projects (such as capital flow, technology sharing, partnership, etc.) serve as edges; GNN (Graph Neural Network) is a deep learning model capable of processing graph data, which analyzes the associations and influences between nodes (projects) through an information propagation mechanism. Specifically, GNN passes the information of a node to its neighbor nodes through the "message passing" process and aggregates information through multiple layers of propagation; this method can uncover potential connections and influences between projects. For example, if there is a problem with the capital chain of a project, GNN can analyze which related projects may be affected and then issue a risk warning; based on the training results of GNN, potential connections between projects are identified. For example, two projects that seemingly have no direct competitive relationship may have an impact due to a certain business cooperation or capital relationship; through this association analysis, entrepreneurs can identify potential cooperation opportunities, market competition situations, and possible resource sharing relationships in advance. Using the GNN analysis results in S42, the risk propagation paths between startup projects are identified. For example, the technical failure, shortage of funds, or market failure of a project may affect other projects through channels such as investors, partners, or the supply chain; the identification of risk propagation paths helps entrepreneurs understand how risks spread between projects and enables them to take measures in advance to avoid or mitigate risks; according to the analysis of the propagation paths, the system automatically generates warning information to remind entrepreneurs to pay attention to potential risks related to their business.For example, if the capital chain of an investor breaks, it may affect multiple related projects. The system will issue a warning to the entrepreneur in advance to help him respond in time. The timeliness and accuracy of risk warnings depend on LSTM's ability to predict future trends and GNN's accurate analysis of project associations, thereby achieving early intervention in potential risks. Warning information is not limited to risk warnings, but may also include response strategy suggestions. For example, on a certain risk propagation path, the system may recommend that certain projects seek additional financial support as soon as possible, or adjust their market strategies; entrepreneurs can optimize their decisions based on these warning information to avoid major losses due to failure to identify risks in a timely manner.

[0106] The effect of the above technical solution is: by inputting the pre-processed data into the trained integrated learning model and combining it with the long short-term memory network (LSTM) architecture, it can effectively capture the long-term dependencies in the time series data and accurately predict the future development trend of the entrepreneurial project. This includes the prediction of key indicators such as user growth, income changes, and market share, so that entrepreneurs can have a clearer understanding of the future potential of the project; using the graph neural network (GNN) to analyze the correlation between entrepreneurial projects can help identify the potential connections and influences between projects. This kind of correlation analysis is not limited to direct competition in the business field, but can also tap into complex implicit connections such as capital flow, technical cooperation, and investor relations. In this way, entrepreneurs can discover external factors that may affect the development of the project in advance, so as to adopt more reasonable response strategies; based on the results of correlation analysis between projects, potential risk transmission paths can be identified. By modeling how risks spread between projects, timely risk warnings can be provided to entrepreneurs. In this way, entrepreneurs can get early warnings before risks occur, take preventive and adjustment measures in advance, effectively reduce losses, and improve risk management capabilities; the technical solution combining LSTM and GNN can not only predict the development trend of the project, but also provide response strategy suggestions based on correlation analysis. Entrepreneurs can optimize their decisions based on the forecast results and risk warning information, formulate more scientific strategies, enhance their market adaptability and risk resistance, and improve their competitive advantages; GNN can be used to analyze the potential connections between entrepreneurial projects to help entrepreneurs explore opportunities for resource sharing, technical cooperation, and market collaboration. This not only helps to improve the overall value of each project, but also provides entrepreneurs with opportunities for cooperation, thereby promoting the coordinated development of projects and optimizing resource allocation; combined with LSTM predictions and GNN correlation analysis, potential growth points and challenges can be identified in a timely manner, and targeted intervention measures can be proposed. In this way, entrepreneurs can make timely adjustments before the challenges are fully revealed to avoid irreversible losses.

[0107] In one embodiment of the present invention, the S41 includes:

[0108] S411. Align all the time series data input into the LSTM model on the time axis and perform normalization processing; based on business understanding and data characteristics, select and construct features that have a significant impact on the prediction results, and use statistical methods and machine learning algorithms to evaluate the feature importance;

[0109] S412. Based on a deep neural network architecture containing multiple LSTM layers, capture the long-term dependencies in the time series data; each layer of LSTM units is responsible for extracting features at different time scales, and gradually deepen the understanding of the time series through the stacking method;

[0110] S413. Incorporate an attention mechanism into the LSTM model to enable the model to dynamically focus on key time periods and features in the time series; tune the hyperparameters of the LSTM model through grid search, and at the same time, use a cross-validation strategy to evaluate the generalization ability of the model;

[0111] S414. Use time series analysis techniques to further interpret the prediction results of the LSTM model, identify the trend components, seasonal components, and residual components of key indicators such as user growth and revenue changes; at the same time, identify outliers in the prediction results through an anomaly detection algorithm to provide early warning information for entrepreneurs;

[0112] S415. Combine the business background and market demand, conduct in-depth analysis of the prediction results, identify potential growth points and challenges of the startup project, and intuitively display the prediction results and analysis conclusions through visualization tools;

[0113] S416. Compare and verify the prediction results of the LSTM model with the actual situation, evaluate the prediction accuracy and reliability of the model; at the same time, collect the feedback and actual needs of entrepreneurs, and iteratively optimize the LSTM model according to the verification results and feedback;

[0114] The working principle of the above technical solution is as follows: Before inputting into the LSTM model, all time series data needs to be preprocessed first. The specific steps are as follows:

[0115] Time alignment: Align the input time series data on the time axis to ensure that the timestamps between the data are consistent. This is because the LSTM needs to process sequence data with time dependence, and misaligned timestamps may affect the model performance.

[0116] Normalization processing: Normalize all the time series data to make them have the same scale (for example, using Z-score normalization). Normalization helps to accelerate the convergence of the model and avoid certain features dominating the model due to different scales.

[0117] Feature Engineering: Based on business understanding and data characteristics, select features that have a significant impact on the prediction results, such as historical growth rate, seasonal fluctuations, market trends, competitor dynamics, etc. These features can be evaluated and extracted through domain knowledge, statistical methods (such as correlation analysis), and machine learning algorithms (such as feature selection algorithms) to ensure that the model can focus on the most important factors for the results.

[0118] The LSTM (Long Short-Term Memory Network) model is designed as a multi-layer deep neural network, and its working principle is as follows:

[0119] Multi-layer LSTM Architecture: LSTM is composed of multiple LSTM units stacked together, and each unit captures long-term dependencies in time series data. LSTM in different layers can extract features at different time scales, and the deep network enables the model to gradually deepen its understanding of the data.

[0120] Hierarchical Structure: The lower LSTM layers handle shorter-term dependencies, while the higher LSTM layers focus on longer-term dependencies. This multi-level structure enables the model to extract information at different time scales, thereby improving the accuracy of prediction.

[0121] An attention mechanism is incorporated into the LSTM model, and its role is to enable the model to dynamically focus on key time periods and features in the time series. The working principle is as follows:

[0122] Adaptive Weights: Through the attention mechanism, the model can assign different weights to different time points in the time series, especially focusing on the moments that are most valuable for the current prediction (such as holidays, major events, etc.).

[0123] Hyperparameter Tuning: By using the Grid Search method, automatically adjust the hyperparameters of the LSTM model (such as the number of LSTM layers, the number of units, the learning rate, the regularization parameter, etc.) to find the optimal model configuration. In addition, evaluate the generalization ability of the model through a cross-validation strategy to ensure that the model not only performs well on the training data but also can effectively predict on unseen data.

[0124] To further enhance the interpretability and prediction reliability of the LSTM model, time series analysis techniques (such as ARIMA, STL decomposition, etc.) are used to post-process the prediction results of the model:

[0125] Trend, Seasonality, and Residual Analysis: Through the ARIMA model or the STL (Seasonal-Trend Decomposition) method, decompose the time series results predicted by LSTM into trend components, seasonal components, and residual components. This analysis helps to identify the basic patterns in the data and improve the understanding of future trends.

[0126] Anomaly Detection: Use anomaly detection algorithms (such as detection methods based on statistics or the Isolation Forest algorithm) to identify outliers in the prediction results. These outliers may represent systematic errors or unexpected fluctuations. Detecting anomalies in a timely manner helps entrepreneurs to warn of potential risks or abnormal events.

[0127] Through in-depth analysis of the prediction results, identify the potential growth points of the project (such as emerging markets, the launch of new product lines) and possible challenges (such as increased competition, policy changes, etc.). This provides clear strategic guidance for entrepreneurs. Use visualization tools such as line charts, bar charts, heat maps, etc. to intuitively display the prediction results and key analysis conclusions. These charts can help decision-makers clearly see trend changes and make more rational and data-supported decisions.

[0128] Compare the prediction results of the LSTM model with the actual data to evaluate the prediction accuracy and reliability of the model. Evaluate the prediction effect by calculating error metrics (such as mean squared error, mean absolute error, etc.). Collect the feedback and actual needs of entrepreneurs and use this feedback information as the basis for iterative optimization. According to the verification results and feedback, adjust the structure of the LSTM model, increase the feature dimension, or introduce new time series analysis methods to improve the accuracy and business adaptability of the model.

[0129] The effects of the above technical solution are as follows: Through the deep learning characteristics of the LSTM model and the introduction of the attention mechanism, the model can effectively capture the long-term dependence relationships in time series data, thereby providing more accurate prediction results. At the same time, the combination of time series analysis techniques (such as ARIMA, STL decomposition) and anomaly detection can further refine the analysis of data trends, seasonality, and abnormal fluctuations, reduce model errors, and improve the reliability of predictions; By adopting hyperparameter tuning (such as grid search) and cross-validation, the model can self-adjust under different configurations to find the optimal combination of hyperparameters. This enables the model to flexibly adapt to different datasets and market changes, thereby maintaining a high prediction accuracy; Through business understanding and data characteristic evaluation, features with significant impacts are selected and constructed, such as historical growth rates, seasonal fluctuations, market trends, etc. This feature engineering can not only improve the learning efficiency of the model but also help entrepreneurs deeply understand the changing trends and influencing factors in the market, providing strong data support for subsequent decision-making; Through visualization tools (such as line charts, bar charts, heat maps, etc.), complex time series data and prediction results are transformed into intuitive information, helping entrepreneurs more easily identify potential growth points and market challenges. Intuitive charts can effectively support business decisions, enabling decision-makers to make more accurate and scientific choices based on data; By using anomaly detection algorithms to timely identify abnormal points in predictions, early warning information can be provided to entrepreneurs, helping them respond more quickly to market changes and avoid potential risks. For example, if certain prediction indicators suddenly deviate from historical trends, the model can automatically capture this, thereby providing timely early warnings for business decisions; After integrating time series analysis techniques and the attention mechanism, the model not only provides prediction results but also gives key influencing factors and changing trends, making the decision-making process of the model more transparent and interpretable. Entrepreneurs can clearly understand which factors have a significant impact on the results, thereby better optimizing strategies; Through comparison and verification with actual data and collection of entrepreneurs' feedback, the model is continuously iteratively optimized. The model is adjusted and optimized according to the actual situation (such as introducing new time series analysis techniques, increasing feature dimensions, etc.), enabling the model to always maintain a close connection with market demands and avoid overfitting or obsolescence; This technical solution has high flexibility and can adapt to different business scenarios and data changes. For example, the LSTM model can process different types of time series data (such as user growth, revenue changes, market share, etc.) according to different business requirements, and meet the prediction needs of different fields through feature selection and hyperparameter adjustment.

[0130] In one embodiment of the present invention, the S412 includes:

[0131] Initialize a neural network layer containing at least one layer of LSTM units as the basic layer of the model; this layer is responsible for initially extracting the temporal features in the time series data;

[0132] On top of the basic LSTM layer, stack more LSTM units layer by layer to form a deep LSTM network; moreover, the output of the previous layer is used as the input for each layer of LSTM units, and by extracting features at different time scales layer by layer, the understanding of time series data is gradually deepened;

[0133] Optimize the time step and batch size of each layer of LSTM units according to the length of the time series data and the computing resources of the model;

[0134] Introduce bidirectional LSTM units into the multi-layer LSTM architecture, enabling the model to consider both the forward and backward information of time series data simultaneously;

[0135] Replace the LSTM units with GRU units in some layers to simplify the model structure and reduce the computational amount, while maintaining the ability to capture long-term dependencies; through comparative experiments, evaluate the performance differences between GRU and LSTM on specific tasks.

[0136] The working principle of the above technical solution is as follows: LSTM (Long Short-Term Memory network) is a special type of Recurrent Neural Network (RNN) that can capture long-term dependencies in time series through "memory cells". In this layer, LSTM receives the input time series data and filters and updates the information through its internal gating mechanism (input gate, forget gate, output gate), thereby extracting the temporal features in the data. The goal of the basic LSTM layer is to initially capture the patterns and trends in the time series data; each layer of LSTM cells receives the output of the previous layer as input and further extracts more complex temporal features based on this input. Each layer of LSTM focuses on capturing the dependencies in the data at different time scales. For example, the lower-layer LSTM may focus on short-term dependencies, while the higher-layer LSTM can capture long-term trends and patterns. Through this layer-by-layer in-depth feature learning, the network can gradually deepen its understanding of the time series; Time step: In time series data, it refers to the time point at which the model inputs each time. The choice of time step will affect the memory ability and computational complexity of the LSTM cell. For longer time series data, a longer time step may be required to capture sufficient historical information; Batch size: It refers to the size of the data batch input into the model during each training. A larger batch size can improve training efficiency, but it may also lead to higher computational resource consumption. An appropriate batch size can find a balance between computational efficiency and memory consumption; The ordinary LSTM model can only process one direction of the time series (from the past to the future). While **Bidirectional LSTM (Bi-LSTM)** builds two LSTM networks simultaneously: one processes the forward information of the time series (from the past to the future), and the other processes the reverse information of the time series (from the future to the past). Bidirectional LSTM can capture more context information in the time series data. Especially in prediction tasks, it can consider both future and past contexts simultaneously, enhancing the expressive ability of the model; GRU (Gated Recurrent Unit) is a simplified version of LSTM. Although both can capture long-term dependencies in time series, the structure of GRU is simpler than that of LSTM. GRU has only two gates (update gate and reset gate), so it is more computationally efficient than LSTM. When replacing LSTM with GRU in some layers, it can reduce computational resource consumption. Especially when the computing power is limited or the data scale is large, GRU can accelerate the training of the model. By comparing the performance of LSTM and GRU in specific tasks, the best unit type can be selected to achieve a better balance between performance and computational efficiency; In the experiment, GRU and LSTM will be respectively applied to the same time series data task to compare their training effects, prediction accuracies, computational efficiencies and other indicators. Generally, LSTM performs better in tasks that require capturing longer time dependencies, while GRU, due to its simple structure and fast computational speed, is suitable for some tasks with high requirements for computational resources but without obvious long-term dependencies.Through the analysis of experimental results, model designers can choose LSTM or GRU according to the characteristics of the task, or use them in combination in some layers.

[0137] The effects of the above technical solutions are as follows: By using the basic LSTM layer and stacking LSTM units layer by layer, the model can better learn and extract temporal features in time series. Each layer of LSTM units can capture features at different time scales, enabling the model to gradually deepen its understanding of the data and improve prediction accuracy; LSTM units are naturally good at handling long-term dependencies, and by deeply stacking LSTM layers, the model can more deeply mine long-term patterns in the data. This can enhance the model's ability to handle complex time series tasks, such as applications in financial market prediction, meteorological data analysis, etc.; By adjusting the time step and batch size of each layer of LSTM units, the training process can be optimized according to the characteristics of time series data and the limitations of computing resources, improving training efficiency and effectively controlling the consumption of computing resources. This optimization is particularly important for large-scale data processing; Introducing bidirectional LSTM units enables the model to not only utilize past information (forward) of time series data but also consider future information (backward). This bidirectional learning can more comprehensively understand the context information of the data and perform excellently in sequence prediction and classification tasks; By using GRU (Gated Recurrent Unit) instead of LSTM units in some layers, the model structure can be simplified and the computational amount can be reduced, thereby improving computational efficiency. GRU has fewer parameters than LSTM, so the calculation is more efficient, and it can still capture long-term dependencies in the data. This optimization can maintain the performance of the model under limited computing resources; By evaluating the performance differences between LSTM and GRU on different tasks through comparative experiments, the model can flexibly select the most suitable unit type according to the actual requirements of the task. For certain specific tasks, GRU may provide better training speed and smaller computational overhead, while for other tasks, the complexity and stronger long-term memory ability of LSTM may be more advantageous; Through the introduction of deeply stacked LSTM networks and bidirectional LSTM, the model can be trained on more complex time series data, thereby enhancing its generalization ability for different types of data. By reasonably replacing LSTM and GRU units, the model performs more stably when processing multiple data sets, helping to solve practical problems in different fields.

[0138] In one embodiment of the present invention, S42 includes:

[0139] Regarding each startup project as a node of the graph, extracting the key attributes of the project as the features of the node, and defining the edges in the graph according to the actual association situation between projects; at the same time, calculating weights for each edge to reflect the strength and importance of the association relationship. The calculation of weights can be based on multiple factors;

[0140] Select a suitable GNN model architecture. In the GNN model, the feature representations of nodes are continuously updated through information propagation and aggregation among nodes.

[0141] Use supervised or unsupervised learning methods to train the GNN model. In supervised learning, construct a training set based on known project association relationships and define a suitable loss function to optimize the model parameters. In unsupervised learning, use graph embedding techniques to learn the low-dimensional representations of nodes, preserving the local and global structural information of the graph.

[0142] Utilize the trained GNN model to calculate the similarity between project nodes and display the association relationships between projects in a graphical manner.

[0143] Based on the node similarity and the visualization results of association relationships, deeply analyze the potential connections and impacts between projects, including identifying groups of projects with similar characteristics or development trajectories, discovering potential cooperation or investment opportunities, and assessing the competitive pressure between projects. And through dynamic graph processing techniques.

[0144] The working principle of the above technical solution is as follows: First, regard each startup project as a node in the graph. Each node contains the key attribute information of the project, and these attributes include:

[0145] Project type: For example, software, hardware, consumer goods, etc.

[0146] Industry: Such as technology, healthcare, finance, etc.

[0147] Founder background: The experience, skills, previous startup projects of the founder, etc.

[0148] Financing round: Such as angel round, Series A, Series B, etc., reflecting the financing stage.

[0149] User scale: The number of users of the project or the market penetration rate.

[0150] Revenue situation: The revenue data or growth trend of the project.

[0151] These key attributes provide rich feature representations for the nodes, helping the model learn and identify the similarity of projects.

[0152] Construct the edges of the graph according to the association relationships between projects (such as investment relationships, cooperation relationships, competitive relationships, etc.). Each edge represents a certain relationship between two projects, which can be:

[0153] Investment relationship: A project is invested in by another project or an investor.

[0154] Cooperation relationship: There is a partnership between two projects, such as joint research and development, market expansion, etc.

[0155] Competitive relationship: Two projects compete in the same market or industry.

[0156] The weight of each edge needs to be calculated according to the strength of the association. The factors for calculating the weight include:

[0157] Investment amount: The larger the investment quantity, the higher the weight of the edge.

[0158] Cooperation frequency: The higher the number or frequency of cooperation, the higher the weight of the edge.

[0159] Market share overlap: The degree of overlap in market share or target user groups between projects.

[0160] To perform efficient feature learning through the graph structure, different GNN model architectures can be selected, such as:

[0161] Graph Convolutional Network (GCN): Updates the representation of nodes through the aggregation of neighbor nodes, suitable for data processing of static graphs.

[0162] Graph Attention Network (GAT): Assigns different weights to different neighbor nodes when propagating information between nodes, suitable for processing graphs with heterogeneous nodes.

[0163] Graph Isomorphism Network (GIN): Captures more complex structural information between nodes through a powerful aggregation function, suitable for tasks that require high sensitivity to structure. In these models, the information propagation and aggregation between nodes are carried out through multi-layer graph neural networks, and finally the feature representation of each node is updated. These feature representations reflect the relationships between nodes and their context information in the entire graph.

[0164] Construct a training set through the known project association relationships. By defining appropriate loss functions (such as cross-entropy loss, mean squared error loss, etc.), the parameters of the model are optimized. The loss function will be optimized based on node similarity and association relationships, with the goal of making similar project nodes close in the embedding space and different category project nodes far apart. If there is no clear labeled data, graph embedding techniques (such as DeepWalk, Node2Vec, etc.) can be used to learn the low-dimensional representations of nodes. In this case, the model will optimize the embedding algorithm to retain the local and global structural information of the graph, ensuring that the embedding representations of project nodes can reflect their potential relationships in the graph. Once the GNN model is trained, the similarity between project nodes can be calculated through the updated node features. Common similarity measurement methods include:

[0165] Cosine similarity: Measures the angular similarity between the feature vectors of two nodes.

[0166] Euclidean distance: Measures the distance of node features in space.

[0167] Based on these similarity measures, the relationship strength between projects can be calculated for further correlation analysis.

[0168] Using the visualization method of graphs, the association relationships between projects are displayed. Through visualization, the investment relationships, cooperation relationships, and competition relationships between projects can be understood more intuitively. According to node similarity, project groups with similar characteristics or development trajectories are discovered; by identifying the similarity or complementarity between projects, potential cooperation or investment opportunities are discovered; by analyzing the competition relationships between projects, the competition pressure of projects in the industry is evaluated. If the dynamic changes of project relationships need to be captured (for example, over time, the investment, cooperation, or competition relationships between projects change), temporal graphs or evolutionary graph models can be used to handle the dynamic evolution of graphs;

[0169] Temporal graph: Each edge is associated with a timestamp, reflecting the change of the relationship between projects over time;

[0170] Evolutionary graph: By tracking the historical changes of the graph, the evolution process of project relationships at different time points is analyzed;

[0171] Through this dynamic graph processing method, the changes of project relationships can be captured dynamically, the entrepreneurial ecosystem at different stages can be analyzed, and predictions can be made according to the time change trend.

[0172] The effects of the above technical solution are as follows: By representing each startup project as a node in a graph and combining the key attributes of the project, it is possible to comprehensively integrate the data characteristics of each project, enabling the multi-dimensional information of each project to be fully reflected in the graph model; Utilizing the advantages of the graph structure, it is possible to establish edges such as actual investment, cooperation, and competition relationships between nodes, capturing the potential complex relationships and influences between projects and providing a basis for subsequent analysis; Through the information propagation and aggregation between nodes in the GNN model, the expression ability of node features can be effectively improved, thereby enhancing the accuracy of project similarity analysis; Whether through supervised learning or unsupervised learning, GNN can optimize the representation of nodes according to the graph structure, enabling similar projects to be close in the embedding space while different types of projects can be correctly distinguished; By calculating the similarity between projects and visual analysis, it is possible to discover which projects have similar characteristics, development trajectories, or market directions, thereby revealing potential partners or investment opportunities; For investors or entrepreneurs, potential cooperation relationships or investment prospects within the industry can be identified through this method; By analyzing the competition relationships between projects, the competitive pressure can be evaluated and the competition pattern among different projects in the market can be insightfully understood; The graph neural network can not only reveal static competition relationships but also capture the time-varying competition situation through dynamic graph technology, helping enterprises and investors predict market trends; Through temporal graph or evolutionary graph processing technology, it is possible to track the evolution of project relationships over time, identify the dynamic changes in project relationships, and predict future relationship trends; This dynamic analysis helps investors, industry analysts, and entrepreneurs predict industry trends and make strategic arrangements in advance; By graphically displaying the association relationships between projects, the relationships and influences between each project can be intuitively presented, helping decision-makers quickly grasp the overall picture of the startup project ecosystem; The visualization results provide effective support for subsequent in-depth analysis, facilitating decision-makers to discover potential opportunities and threats; Through efficient graph neural network analysis, the quality and precision of decision-making can be improved, reducing the bias of human judgment; In terms of risk assessment, potential investment risks, competition risks, or market risks can be discovered through the graph model, providing data support and decision-making basis for risk management.

[0173] In one embodiment of the present invention, the S5 includes:

[0174] S51. Collect relevant data of students; Based on the collected data, construct a student user profile, where the user profile includes the students' hobbies, learning styles, and project requirements;

[0175] S52. According to the characteristics of the user profile, divide the students into different groups; Establish a resource library containing teaching resources such as courses, cases, and expert lectures, and based on the user profile and the characteristics of the student groups, screen out suitable teaching resources for the students from the resource library according to the recommendation algorithm.

[0176] S53. Generate personalized teaching suggestions based on the results of the recommendation algorithm, and push the suggestions to students through various methods such as email, text message, and APP push.

[0177] S54. Based on the A / B test plan, divide students into an experimental group and a control group, conduct experiments using different teaching strategies, collect the learning performance data of students in the experimental group and the control group, and conduct comparative analysis.

[0178] S55. Evaluate the effectiveness of the teaching strategy based on the comparative analysis results, and optimize and adjust the teaching strategy based on data-driven decision support.

[0179] The working principle of the above technical solution is as follows: First, collect relevant data of students through various channels. The data content includes but is not limited to the basic information of students (such as age, grade, major, etc.), learning history (including course grades, learning time, participation in learning activities, etc.), and project participation (such as practical activities, competitions, extracurricular projects, etc.). These data can provide the learning situation and behavioral characteristics of students. By sorting and analyzing the collected data, a personalized user profile of each student is constructed. The profile not only includes the basic information of the student but also:

[0180] Hobbies: Identify the student's interest areas by analyzing the projects, elective courses, extracurricular activities, etc. participated by the student.

[0181] Learning style: Understand the student's learning preferences by analyzing the student's learning methods (such as preferring self-study, cooperative learning, visual / audio, etc.).

[0182] Project requirements: Based on the student's academic background and career goals, identify their requirements in a specific project, such as the need for a certain skill or knowledge.

[0183] Based on the constructed user profiles, use clustering analysis or other classification algorithms to subdivide students into different groups. For example, grouping can be done according to dimensions such as learning styles, interests, and academic levels. The characteristics of each student group will help determine their needs for different teaching resources; Integrate various teaching resources (such as courses, cases, expert lectures, online lectures, extracurricular activities, etc.), and classify and label the resources to match the needs of the student groups. The information contained in the resource library should be able to meet the learning needs and interests of different groups; Use recommendation algorithms (such as content-based recommendation, collaborative filtering, deep learning, etc.) to screen out the most suitable teaching resources from the resource library for the student groups. The recommendation algorithm will push the most relevant courses, cases, lectures, etc. based on the characteristics of the student groups; Based on the results of the recommendation algorithm, generate personalized teaching suggestions for each student or student group. The suggestions can include recommended courses, projects, learning paths, or skill improvement directions, etc., aiming to help students achieve learning goals more effectively; Communicate the teaching suggestions to students through different communication channels (such as emails, text messages, APP push, etc.). Select the appropriate method for pushing according to the preferences of the students to ensure that the suggestions can reach the students in a timely manner and improve the students' learning engagement; Based on the design method of A / B testing, divide the students into an experimental group and a control group. The experimental group adopts specific teaching strategies, while the control group uses traditional teaching methods. Through this grouped experiment, the effects of different teaching strategies can be tested; Collect the learning data generated by the experimental group and the control group during the experiment, including indicators such as students' academic performance, engagement, and learning time. Evaluate the advantages and disadvantages of the teaching strategies by comparing and analyzing the two groups of data; Based on the comparative analysis results of A / B testing, evaluate the impact of different teaching strategies on students' learning effects. For example, whether a certain strategy significantly improves students' grades, engagement, learning satisfaction, etc.; According to the data analysis results, combined with educational psychology and teaching theories, optimize and adjust the teaching strategies. The teaching team can rely on a data-driven decision support system to implement different teaching methods for different groups of students; Continuously adjust the teaching strategies according to the feedback data. For example, if a certain teaching strategy shows good results in the experimental group, it can be promoted among all students. Conversely, for strategies with poor effects, they can be optimized or replaced.

[0184] The effects of the above technical solution are as follows: By collecting data such as students' basic information, learning history, interests and hobbies, and learning styles, and constructing user portraits based on this data, it is possible to accurately identify students' needs, thereby providing customized teaching resources. Such personalized recommendations can help students better obtain learning materials that match their interests and needs, improving learning efficiency and effectiveness; By segmenting students and pushing corresponding teaching resources according to the characteristics of the segmented groups, it is possible to ensure the effective matching of teaching resources, avoiding the problems of single and generalized traditional teaching resource pushing methods. Students can receive resources that are more in line with their personal needs and goals, enhancing the learning experience; By pushing personalized teaching suggestions through multiple channels (such as email, SMS, APP push, etc.), it is possible to increase students' attention and participation in teaching resources and learning suggestions. The personalized pushing method can also be adjusted according to students' preferences, enhancing the initiative and satisfaction of learning; Using experimental design schemes such as A / B testing to verify different teaching strategies can collect a large amount of learning performance data and evaluate the effectiveness of different strategies through comparative analysis. This data-driven decision support helps to continuously optimize teaching strategies, thereby improving teaching quality and students' learning outcomes; According to the comparative analysis results of the learning performance of the experimental group and the control group, it is possible to clearly see which teaching strategies are effective and which may need to be optimized or adjusted. The data-driven approach enables real-time adjustment of teaching strategies and continuous improvement based on students' feedback, thereby achieving long-term improvement in teaching quality; Through continuous personalized recommendations and strategy optimization, the educational process can be made more accurate and efficient, and students can learn in an environment that suits their interests and learning styles, thereby enhancing the overall educational quality and learning effect, and ultimately helping students achieve better academic performance and career development; The establishment of a resource library and the intelligent screening of teaching resources can ensure the maximum utilization of teaching resources. By analyzing the needs of the student group, waste of resources can be avoided, and resource allocation can be made more reasonable, improving the use efficiency of educational resources.

[0185] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. An education method based on multi-dimensional data fusion, characterized in that, The method includes: S1. Obtain the data source; S2. Extract relevant features, form a feature vector, and perform dimensionality reduction and fusion; S3. Construct an ensemble learning model; S4. Through a long short-term memory network architecture, combined with time series analysis, predict the future development trend of the startup project, and use a graph neural network to analyze the correlation relationship between startup projects to identify potential risk propagation paths; S5. Generate personalized teaching resources and suggestions.

2. The educational method based on multi-dimensional data fusion according to claim 1, characterized in that, The S1 includes: S11. Determine the data collection method; S12. Preprocess the collected raw data and extract key information; S13. Integrate the preprocessed data into a unified data warehouse.

3. The educational method based on multi-dimensional data fusion according to claim 1, characterized in that, The S2 includes: S21. Extract relevant features from the raw data and screen the extracted features; S22. Construct a feature vector and perform normalization on the feature vector; S23. Construct a dimensionality reduction and fusion model for the feature vector; S24. Perform dimensionality reduction and fusion on the feature vector to form a low-dimensional representation vector.

4. The educational method based on multi-dimensional data fusion according to claim 3, characterized in that, The S23 includes: Construct an autoencoder with multiple hidden layers, and each layer undertakes different feature extraction and dimensionality reduction tasks; by stacking multiple encoding and decoding layers, gradually abstract high-level feature representations; Introduce a sparsity constraint in the encoding layer; introduce cross-layer connections between the hidden layers of the autoencoder to allow direct interaction of feature information at different levels; introduce an attention mechanism in the decoding layer to dynamically adjust the contribution of different features during the reconstruction process; Use grid search to finely tune the hyperparameters of the autoencoder; Introduce an L2 regularization term in the loss function. At the same time, by adding a Dropout layer, randomly discard some neurons during the training process; and monitor the loss or accuracy metrics on the validation set during the training process. When the metrics no longer improve, stop training in advance; Add noise to the input data to evaluate the dimensionality reduction effect of the autoencoder model; According to the evaluation results, adjust the relevant parameters of the autoencoder to form a closed-loop optimization process.

5. The educational method based on multi-dimensional data fusion according to claim 1, wherein, The S3 includes: S31. Determine the base learners; combine multiple base learners into an ensemble learning model; S32. Divide the preprocessed data into a training set, a validation set, and a test set; S33. Iteratively optimize the model according to the validation results; S34. Adjust the variable weights in the ensemble learning model.

6. The educational method based on multi-dimensional data fusion according to claim 1, wherein The S4 includes: S41. Input the preprocessed data into the trained ensemble learning model. Through a long short-term memory network architecture, combined with time series analysis, predict the future development trend of the startup project; analyze the prediction results to identify potential growth points and challenges; S42. Use a graph neural network to analyze the correlation relationship between startup projects to identify potential connections and influences between projects; S43. Based on the project correlation analysis results, identify potential risk propagation paths, and provide early warning information for entrepreneurs according to the risk propagation paths.

7. The educational method based on multi-dimensional data fusion according to claim 6, characterized in that, The S41 includes: S411. Align all the time series data input into the LSTM model on the time axis and perform normalization processing; Based on business understanding and data characteristics, select and construct features that have a significant impact on the prediction results, and use statistical methods and machine learning algorithms to evaluate the feature importance; S412. Based on a deep neural network architecture containing multiple LSTM layers, capture the long-term dependencies in the time series data; Each layer of LSTM units is responsible for extracting features at different time scales, and gradually deepen the understanding of the time series through stacking; S413. Incorporate an attention mechanism into the LSTM model to enable the model to dynamically focus on key periods and features in the time series; Tune the hyperparameters of the LSTM model through grid search. At the same time, adopt a cross-validation strategy to evaluate the generalization ability of the model; S414. Use time series analysis techniques to further interpret the prediction results of the LSTM model, identify the trend components, seasonal components, and residual components of key indicators; At the same time, identify outliers in the prediction results through anomaly detection algorithms to provide early warning information for entrepreneurs; S415. Combine the business background and market demand, conduct in-depth analysis of the prediction results, identify potential growth points and challenges of the startup project, and intuitively display the prediction results and analysis conclusions through visualization tools; S416. Compare and verify the prediction results of the LSTM model with the actual situation, evaluate the prediction accuracy and reliability of the model; At the same time, collect the feedback and actual needs of entrepreneurs, and iteratively optimize the LSTM model according to the verification results and feedback; 8. The educational method based on multi-dimensional data fusion according to claim 6, characterized in that, The above S412 includes: Initialize a neural network layer containing at least one layer of LSTM units as the basic layer of the model; This layer is responsible for initially extracting the temporal features in the time series data; On top of the basic LSTM layer, stack more LSTM units layer by layer to form a deep LSTM network; Moreover, each layer of LSTM units takes the output of the previous layer as input, and gradually deepen the understanding of the time series data by extracting features at different time scales layer by layer; Optimize the time step and batch size of each layer of LSTM units according to the length of the time series data and the computing resources of the model; Introduce bidirectional LSTM units in the multi-layer LSTM architecture, enabling the model to consider both the forward and backward information of the time series data simultaneously; Replace the LSTM units with GRU units in some layers to simplify the model structure and reduce the computational amount, while maintaining the ability to capture long-term dependencies; Through comparative experiments, evaluate the performance differences between GRU and LSTM on specific tasks.

9. The educational method based on multi-dimensional data fusion according to claim 1, characterized in that The above S5 includes: S51. Build a student user profile; S52. Screen out teaching resources suitable for students from the resource library; S53. Push suggestions to students in various ways; S54. Divide students into an experimental group and a control group for comparative analysis; S55. Optimize and adjust the teaching strategy.

10. An education platform based on multi-dimensional data fusion, characterized in that, It includes a memory, a processor, and a computer program stored on and executable on the memory, and the processor executes the program to implement the education method based on multi-dimensional data fusion according to any one of claims 1-9.

Images