AI-based investment project feasibility intelligent analysis and decision support system

Abstract

The invention relates to the field of finance, and discloses an AI-based investment project feasibility intelligent analysis and decision support system, which comprises a multi-modal data fusion engine used for extracting a feature vector of an unstructured text through a BERT variant model, carrying out joint coding on the feature vector and a structured data feature, and outputting fused multi-modal data; the dynamic knowledge graph construction module is connected with the multi-modal data fusion engine; the deep reinforcement learning decision module comprises a feasibility prediction network and a risk assessment network which are cooperatively trained; an interpretable AI module; and a closed loop feedback module. By means of a multi-modal data fusion engine and a deep reinforcement learning decision module, automatic processing and intelligent analysis of data are achieved, and the analysis time of a single project is greatly shortened to 8 minutes. The high-efficiency processing speed enables the investment institution to quickly respond to the market change, timely grasp the investment opportunity, and occupy the decision-making precedence in the financial market which changes instantaneously.

Description

Technical Field

[0001] The present invention relates to the field of financial technology, and in particular to an AI-based intelligent feasibility analysis and decision support system for investment projects. Background Art

[0002] In the field of investment project evaluation, traditional methods rely primarily on manual empirical analysis, which presents numerous drawbacks. Firstly, the evaluation process is subject to significant subjective bias, and differences in experience and understanding among evaluators can lead to inconsistent and objective results. Secondly, manual evaluation is inefficient, unable to rapidly process large amounts of complex data, and unable to meet the rapidly changing demands of the market. Furthermore, traditional evaluation methods are weak in their ability to integrate multi-dimensional data, making it difficult to comprehensively consider all factors influencing investment projects.

[0003] At the same time, existing AI models also have obvious shortcomings in investment project evaluation applications. Most AI models only analyze a single data source, such as focusing only on financial data, while ignoring unstructured data such as policy texts, industrial chain relationships, and public opinion. These unstructured data contain rich information and have a significant impact on the feasibility of investment projects. The lack of dynamic correlation analysis between them makes the evaluation results inaccurate and incomplete.

[0004] At the same time, existing AI models lack interpretability, rendering their decision-making processes like a "black box." This makes it difficult for investment institutions to understand the underlying rationale behind these models, leading to low trust in AI decisions. This, to a certain extent, limits the widespread application of AI technology in the investment sector. Furthermore, existing systems lack a real-time feedback loop, making it impossible to update evaluation models in response to market dynamics, making it difficult to ensure the model's adaptability and accuracy. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an AI-based intelligent analysis and decision support system for investment project feasibility, which solves the problem that AI models also have obvious shortcomings in investment project evaluation applications.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: an AI-based investment project feasibility intelligent analysis and decision support system, comprising:

[0007] Multimodal data fusion engine, which is used to extract feature vectors of unstructured text using a BERT variant model, jointly encode them with structured data features, and output fused multimodal data;

[0008] A dynamic knowledge graph construction module, connected to the multimodal data fusion engine, uses a graph neural network to construct an industry graph containing enterprise nodes and upstream and downstream relationships, and dynamically updates node features through the industry volatility index;

[0009] A deep reinforcement learning decision module, including a collaboratively trained feasibility prediction network and a risk assessment network, receives the multimodal data and the dynamic knowledge graph, and outputs a feasibility score and a risk assessment result;

[0010] Explainable AI module, which performs feature attribution analysis on decision results based on SHAP values ​​and generates audit trail logs that comply with SEC and FCA regulatory requirements;

[0011] The closed-loop feedback module deploys an online learning mechanism and an industry risk warning signal system, triggering T+1 level model updates based on actual investment deviations and market fluctuations.

[0012] Preferably, the multimodal data fusion engine performs context-sensitive text feature extraction on policy documents and news public opinion through a BERT variant model, uses layer normalization and attention mechanism to generate word feature vectors, performs Z-score normalization on financial data and supply chain data in structured data, and realizes joint encoding through feature vector splicing or weighted fusion formula, where the weight parameters are dynamically adjusted through model training.

[0013] Preferably, the risk keyword quantification method of the BERT variant model includes: identifying a set of policy risk keywords, calculating a weighted sum based on the appearance position weight and the context importance weight; and mapping the word feature vector to a risk score through a trained small classifier.

[0014] Preferably, the node feature update formula of the graph neural network in the dynamic knowledge graph construction module is: in, For the The weight matrix of the layer, is the bias vector, is the activation function.

[0015] Preferably, the industry volatility index calculation includes: calculating the financial volatility index and the market volatility index of the enterprise node set in the industry, and determining the industry volatility index through a weighted formula, where is the industry status weight of node i.

[0016] Preferably, in the deep reinforcement learning decision module, the feasibility prediction network uses temporal difference learning to update the Q-value function, and the update formula is:

[0017] in, are DQN network parameters, is the target network parameter, is the learning rate, is the discount factor.

[0018] Preferably, the deep reinforcement learning decision module is embedded in Monte Carlo tree search, the confidence upper bound formula is adopted in the node selection phase, the investment return distribution is generated based on the policy network in the simulation phase, and the number of node visits and the average reward value are updated through back propagation.

[0019] Preferably, in the explainable AI module, the SHAP value calculation is solved by an approximate algorithm, where the baseline value is the predicted mean of the training set, and the visual display includes a SHAP value heat map of risk keywords and a contribution bar chart of structured features.

[0020] An AI-based intelligent feasibility analysis and decision support method for investment projects, comprising the following steps:

[0021] S1. Data collection and preprocessing: Collect structured and unstructured data, and perform cleaning, standardization, word segmentation, and denoising preprocessing respectively;

[0022] S2. Multimodal Data Fusion: Utilize the BERT variant model to extract unstructured data feature vectors, and then concatenate or weightedly fuse them with structured data feature vectors to generate multimodal fused data vectors.

[0023] S3. Dynamic Knowledge Graph Construction: Based on multimodal fusion data, we use graph neural networks to build an industry upstream and downstream association graph, and update node features and connection weights based on market data fluctuations.

[0024] S4. Investment Project Evaluation: Multimodal fusion data and dynamic knowledge graphs are input into the deep reinforcement learning decision-making module. Feasibility is predicted through the feasibility prediction network, and risk is assessed through the risk assessment network. An evaluation report is generated based on the Monte Carlo tree search simulation results.

[0025] S5. Decision interpretation and compliance checking: Visualize key decision-making basis through the SHAP value attribution system and generate audit trail logs according to regulatory requirements;

[0026] S6. Model update and optimization: Based on the deviation between actual investment results and predicted results, the model parameters are adjusted through the online learning mechanism, and the T+1 model update is triggered in conjunction with industry risk warnings.

[0027] Preferably, in step S2, the text sequence is subjected to layer normalization and multi-layer attention mechanism processing to generate a text feature vector, and the risk score is calculated based on the keyword position and context weight, and the score is quantified and converted into a feature vector.

[0028] The present invention provides an AI-based intelligent feasibility analysis and decision support system for investment projects. It has the following beneficial effects:

[0029] 1. This invention leverages a multimodal data fusion engine and a deep reinforcement learning decision module to achieve automated data processing and intelligent analysis, significantly reducing single-project analysis time to just 8 minutes. This efficient processing speed enables investment institutions to quickly respond to market changes, seize investment opportunities promptly, and gain a proactive decision-making advantage in the ever-changing financial market.

[0030] 2. This invention acquires comprehensive information through multimodal data fusion, accurately depicts industry relationships and risks through dynamic knowledge graphs, and has powerful learning and prediction capabilities through deep reinforcement learning decision modules. The three work together to more accurately assess the potential value of investment projects, provide solid and reliable data support for investment decisions, and reduce the uncertainty of investment decisions.

[0031] 3. The present invention uses a dynamic knowledge graph to update the risk transmission path of the industrial chain in real time, combined with the calculation of the industry volatility index, to effectively identify various risks, help investment institutions formulate response strategies in advance, and reduce risk losses; on the other hand, it automatically generates an explainable report that complies with the ISO 55001 standard, and the impact factor attribution system based on the SHAP value visualizes the decision basis. The regulatory compliance inspection layer automatically generates an audit tracking log to meet the strict regulatory requirements of the financial industry, improve the transparency and compliance of investment decisions, and enhance investment institutions' trust in AI decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a perspective view of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] Example:

[0035] Please see the attached Figure 1, an embodiment of the present invention provides an AI-based intelligent analysis and decision support system for investment project feasibility, including a multimodal data fusion engine, which is used to realize the joint encoding of structured data and unstructured data. Structured data includes financial data, supply chain data, etc., and unstructured data covers policy documents, news and public opinion, etc. A context-sensitive text feature extraction module based on a BERT variant model is used to process unstructured text data such as policy documents and news and public opinion, which can accurately identify policy risk keywords and implement quantitative scoring of these keywords through a specific algorithm, converting unstructured data into feature vectors that can be used for analysis and fusing them with structured data;

[0036] First, the input policy documents or news and public opinion text sequence Feed it into the BERT variant model. In each layer of the model, for each word , first take the feature vector of the previous layer Perform layer normalization operation, namely LayerNorm ,Stabilize network training and make feature distribution more stable.

[0037] By formula Attention softmax Get the attention output, and then use the attention mechanism Attention Calculate the current word The degree of association with other words in the text. Mapped to query vectors , all words Mapping to key vector ,Then, the layer normalization result is added to the attention output to get the current layer word The eigenvector of LayerNorm Attention Repeat the above operations through the multi-layer Transformer architecture to finally obtain the text feature vector Finally, for the policy risk keyword set , traverse each keyword in the text The location of the appearance , determine the weight based on factors such as its location, context importance, etc. , and calculate the word feature vector based on the trained small classifier or regression model Risk Score , all occurrences of Score Sum and get keywords Quantitative score of Score in, Keywords In the text The weight of the first occurrence is determined based on factors such as its location and contextual importance; Based on word feature vector The calculated risk score can be obtained by training a small classifier or regression model.

[0038] For structured data, standardized data interfaces are used for collection, and financial data, supply chain data, and other data are stored and managed in a unified data format. During the data preprocessing phase, data is cleaned to remove duplicate and erroneous data records, and the data is standardized to have a unified dimension and format.

[0039] For unstructured data, web crawler technology is used to collect text data such as policy documents and news and public opinion from data sources such as government websites and news media websites. A context-sensitive text feature extraction module based on a BERT variant model is used to perform word segmentation and vectorization on the collected text data. The BERT variant model is trained to identify policy risk keywords and, based on the contextual information of the keywords in the text, calculates a quantitative score for each keyword, converting the text data into feature vectors. Finally, the feature vectors of the structured data and the feature vectors of the unstructured data are concatenated or fused to achieve joint encoding of multimodal data.

[0040] Suppose the structured data feature vector is , the feature vector of unstructured data extracted by the BERT variant model is The feature vector after joint encoding It can be obtained by splicing or weighted fusion. Splicing method: Weighted fusion method: ,in is a weight parameter learned through model training and used to balance the importance of structured data and unstructured data.

[0041] This multimodal data fusion connects to a dynamic knowledge graph construction module, which uses a graph neural network (GNN) to construct an industry upstream and downstream relationship graph. First, the graph's nodes and edges are determined. Nodes can represent entities such as companies and industries, while edges represent relationships between entities, such as the relationship between suppliers and buyers, or the relationship between upstream and downstream industries. By collecting basic company information and business transaction data, initial connections between nodes are established.

[0042] First, determine the nodes in the industry upstream and downstream relationship map The initial eigenvector of and its neighbor node set .

[0043] Then, in the first When calculating the layer, for the neighbor node set Each node in , its first The feature vector of the layer Multiply by The weight matrix of the layer , and add the bias vector ,get .

[0044] Then, the activation function is performed on the above results (such as ReLU) operation, introducing nonlinearity, and obtaining .

[0045] Finally, the activated result is combined with the node Self The feature vector of the layer Add together to get the node In the The feature vector of the layer , complete the information dissemination and feature update between nodes. Among them, For the The weight matrix of the layer, is the bias vector, is the activation function (such as ReLU). This formula represents the node In the The characteristics of a layer are obtained by weighted aggregation and nonlinear transformation of the characteristics of its neighboring nodes in the previous layer, and then adding them to the characteristics of its own previous layer, thereby realizing the dissemination and fusion of information between nodes and reflecting the impact of industrial chain relationships on node characteristics.

[0046] Real-time market data collection, including updated corporate financial statements, industry news, and policy changes, is used. Based on this data, a graph neural network algorithm is used to update the upstream and downstream industry linkage map, adjusting the connection weights between nodes to reflect changes in the risk transmission path of the supply chain. Furthermore, the industry volatility index calculation submodule analyzes market data fluctuations, embeds graph nodes, and incorporates market sensitivity information into node features, enabling the construction and updating of a dynamic knowledge graph.

[0047] First, determine the set of enterprise nodes in the industry , and calculate each node Financial volatility indicators (such as the standard deviation of revenue growth rates) and market volatility indicators (Such as stock price volatility). Then, based on the company's position in the industry, scale and other factors, determine each node The weight of financial volatility indicators and market volatility index weights . Then, calculate the weighted sum of financial volatility indicators Weighted sum of market volatility indicators , and add them together to get the numerator ;Calculate the sum of all weights at the same time As the denominator. Finally, through the formula Calculate the industry volatility index , used to update the market sensitivity information of graph nodes.

[0048] The knowledge graph construction module is connected to the deep reinforcement learning decision-making module, which uses a dual-network architecture to construct a feasibility prediction network (DQN) and a risk assessment network (PPO). The DQN network is used to predict the feasibility of investment projects. By learning from historical investment project data and market data, it establishes a mapping relationship between state, action, and reward, and predicts the feasibility of projects under different investment decisions. The PPO network is used to assess the risk of investment projects, considering various risk factors such as market risk, policy risk, and financial risk, and outputs the project risk assessment results.

[0049] In DQN, let the state space be , the action space is In state Next, execute the action After that, the reward is , and transfer to the new state Q-value function To evaluate the status Next action The long-term value is updated by the following formula:

[0050]

[0051] in, are DQN network parameters, The target network parameters (regularly copied and updated from the main network), is the learning rate, is the discount factor (used to weigh the importance of current rewards and future rewards). This formula is continuously adjusted by temporal difference learning (TD-learning). Value function, which makes it more accurate in predicting the feasibility value of investment projects under different states and actions;

[0052] The two networks are trained using a collaborative training mechanism. During the training process, multimodal fusion data and a dynamic knowledge graph are used as input. The two networks collaborate with each other, sharing some parameters and information. By continuously adjusting network parameters, the networks can more accurately predict project feasibility and assess risks.

[0053] The PPO algorithm optimizes the policy network by maximizing the objective function , the objective function is:

[0054]

[0055] in, is the importance sampling ratio, which is used to compare the probability of actions under the new and old strategies; is the estimated value of the advantage function, which is used to measure the superiority of the current action relative to the average action; This is a truncation parameter used to limit the magnitude of policy updates and prevent excessive policy updates from causing performance degradation. By optimizing this objective function, the PPO network can learn a better risk assessment strategy and accurately assess the various risks faced by investment projects.

[0056] At the same time, Monte Carlo Tree Search (MCTS) is embedded. In the decision-making process, by simulating a large number of different market scenarios, the investment return distribution under each scenario is calculated to provide a more comprehensive reference for investment decisions. First, the search tree is initialized, and the root node represents the current market status. Then, in the tree node selection stage, for each node , according to the formula

[0057]

[0058] Calculate its upper confidence limit ,in For nodes The average cumulative reward of For nodes Number of visits, is the number of visits to the parent node, For the exploration coefficient, select The node with the largest value is expanded until a leaf node is reached.

[0059] Subsequently, in the leaf node simulation phase, starting from the leaf node, simulation is performed according to a certain strategy (such as a random strategy or a strategy based on the current strategy network) until the termination condition is reached (such as the end of the simulation or the maximum number of simulation steps is reached), and the simulated investment return is obtained. The return is backpropagated to each node of the search tree, and the number of node visits and the average cumulative reward are updated. Finally, the node selection, expansion, simulation and backpropagation steps are repeated to continuously construct the search tree, estimate the investment return distribution under different market scenarios, and provide a basis for investment decisions.

[0060] The deep reinforcement learning decision module is connected to the explainable AI module, which builds an influencing factor attribution system based on SHAP value. After the model training is completed, the SHAP value algorithm is used to calculate the impact of each feature on the evaluation results of each investment project, that is, the SHAP value. These SHAP values ​​are visualized, such as through bar charts, heat maps, etc., to intuitively show which factors have a greater impact on the feasibility and risk assessment results of the investment project, to explain the decision-making basis of the model to investment institutions, and to help the model prediction results and the input feature vector ,SHAP value To measure characteristics The contribution to the prediction results is obtained by solving the following equation:

[0061]

[0062] in, is the baseline value (usually the average of all sample predictions). In actual calculations, approximate algorithms, such as KernelSHAP, are used to sample different feature subsets and calculate the marginal contribution of each feature to the prediction result, thereby obtaining the SHAP value of each feature. This is used to visualize the basis for key decisions.

[0063] A regulatory compliance layer is implemented to review the model's decision-making process and results in accordance with relevant regulatory requirements, including those of the SEC and FCA. Automatically record the model's input data, intermediate calculations, and outputs, generating an audit trail to ensure regulatory compliance and provide a clear basis for regulatory review.

[0064] The explainable AI module is connected to a closed-loop feedback module. This module deploys an online learning mechanism to promptly collect data on actual investment results, including actual returns, costs, and risk, after the implementation of actual investment projects. This module compares the actual investment results with the model's predictions, calculating a deviation. Based on this deviation, optimization algorithms such as gradient descent are used to adjust the parameters of the deep reinforcement learning decision-making module, enabling the model to more accurately predict project feasibility and risk in subsequent evaluations.

[0065] Establish an industry risk early warning signal system to monitor market dynamics and industry data in real time, such as changes in industry indices, changes in policies and regulations, and major news events. Based on pre-set risk indicators and thresholds, the system determines the industry risk level and displays it through different signal light colors (e.g., green for low risk, yellow for medium risk, and red for high risk). Changes in risk levels trigger iterative model updates, achieving a T+1 model update cycle to ensure the model can adapt promptly to market changes.

[0066] A method for implementing an AI-based investment project feasibility intelligent analysis and decision support system includes the following steps:

[0067] S1. Data Collection and Preprocessing

[0068] S101. Establish data collection channels. Collect structured and unstructured data related to investment projects through collaboration with financial data providers and web crawlers. Structured data, such as financial data, is obtained from the company's financial statements and financial databases; supply chain data is collected from the company's supply chain management system. For unstructured data, web crawlers are used to collect policy documents from government websites and news and public opinion data from news websites and social media platforms.

[0069] S102. Clean the collected structured data to remove duplicate records, records with excessive missing values, and erroneous data. Use standardization methods, such as Z-score standardization, to normalize the data to uniform dimensions and distribution. For unstructured data, perform preprocessing operations such as word segmentation, stop word removal, and noise removal to provide a high-quality data foundation for subsequent data fusion and analysis.

[0070] S2. Multimodal Data Fusion

[0071] S202. Input the preprocessed structured and unstructured data into the multimodal data fusion engine. Using a context-sensitive text feature extraction module based on a BERT variant model, feature extraction is performed on unstructured text data such as policy documents and news and public opinion, generating text feature vectors. Simultaneously, feature engineering is performed on the structured data to extract key features.

[0072] S203. Using a data fusion algorithm, the feature vectors of the structured data and the feature vectors of the unstructured data are fused. Methods such as concatenation and weighted summation can be used to integrate the two types of data features into a multimodal fused data vector, which serves as input for subsequent analysis.

[0073] S3 dynamic knowledge graph construction

[0074] S301. Based on multimodal fusion data, use a graph neural network (GNN) to construct an industry upstream and downstream relationship graph. Determine the definitions of the graph's nodes and edges, and initialize the graph structure based on information such as corporate relationships and industry associations in the data.

[0075] S302. Utilize the industry volatility index calculation submodule to analyze market data fluctuations, such as changes in industry sales and company stock price fluctuations. Based on the analysis results, graph nodes are embedded and their feature vectors are updated to reflect market sensitivity. Simultaneously, based on real-time market data, the connections and weights between nodes in the graph are dynamically updated to construct a dynamic knowledge graph that reflects dynamic changes across industries and risk transmission pathways.

[0076] S4. Investment Project Evaluation

[0077] S401. Input the multimodal fusion data and dynamic knowledge graph into the deep reinforcement learning decision module. The feasibility prediction network (DQN) uses the input data to predict the feasibility of the investment project under different decision conditions and outputs a feasibility score. The risk assessment network (PPO) analyzes the various risk factors of the investment project, assesses the project's risk level, and outputs the risk assessment results.

[0078] S402. Using Monte Carlo Tree Search (MCTS), we simulate the investment process under different market scenarios and calculate the distribution of investment returns. By integrating the results of the feasibility prediction network and risk assessment network, along with the investment return distribution derived from the Monte Carlo Tree Search, we produce a feasibility analysis and risk assessment report for the investment project, providing a comprehensive reference for investment decisions.

[0079] S5. Interpretation of decision results and compliance checks

[0080] S501. Analyze investment project evaluation results using the SHAP value impact factor attribution system within the explainable AI module. Calculate the SHAP value for each feature and use visualization tools to display key decision-making factors, enabling investment institutions to understand the model's underlying reasons for making decisions.

[0081] S502. Conduct compliance checks on the investment project evaluation process and results through the regulatory compliance check layer. Automatically generate an audit trail log that complies with SEC / FCA and other regulatory requirements, recording data input, model calculations, and outputs during the evaluation process. This ensures that the investment decision-making process complies with regulatory requirements and improves the transparency and credibility of decision-making.

[0082] S6. Model update and optimization

[0083] S601. After the investment project is implemented, actual investment results are collected and compared with the model prediction results to calculate the deviation value. The deviation value is used as a feedback signal and input into the closed-loop feedback system.

[0084] S602. Utilize online learning mechanisms to adjust the parameters of the deep reinforcement learning decision module based on the deviation value. Optimization algorithms, such as stochastic gradient descent, are used to update the model's weights, enabling it to better adapt to actual market conditions. Simultaneously, based on market dynamics and industry risk changes detected by the industry risk early warning system, iterative model updates are triggered, achieving T+1 model optimization and continuously improving the model's predictive accuracy and adaptability.

[0085] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based investment project feasibility intelligent analysis and decision support system, characterized by: include: Multimodal data fusion engine, which is used to extract feature vectors of unstructured text using a BERT variant model, jointly encode them with structured data features, and output fused multimodal data; A dynamic knowledge graph construction module, connected to the multimodal data fusion engine, uses a graph neural network to construct an industry graph containing enterprise nodes and upstream and downstream relationships, and dynamically updates node features through the industry volatility index; A deep reinforcement learning decision module, including a collaboratively trained feasibility prediction network and a risk assessment network, receives the multimodal data and the dynamic knowledge graph, and outputs a feasibility score and a risk assessment result; Explainable AI module, which performs feature attribution analysis on decision results based on SHAP values ​​and generates audit trail logs that comply with SEC and FCA regulatory requirements; The closed-loop feedback module deploys an online learning mechanism and an industry risk warning signal system, triggering T+1 level model updates based on actual investment deviations and market fluctuations.

2. The AI-based investment project feasibility intelligent analysis and decision support system according to claim 1 is characterized in that: The multimodal data fusion engine uses a BERT variant model to extract context-sensitive text features from policy documents and news and public opinion, adopts layer normalization and attention mechanism to generate word feature vectors, performs Z-score normalization on financial data and supply chain data in structured data, and realizes joint encoding through feature vector splicing or weighted fusion formula, where the weight parameters are dynamically adjusted through model training.

3. The AI-based investment project feasibility intelligent analysis and decision support system according to claim 2 is characterized in that: The risk keyword quantification method of the BERT variant model includes: identifying a set of policy risk keywords, calculating a weighted sum based on the appearance position weight and the context importance weight; and mapping the word feature vector to a risk score through a trained small classifier.

4. The AI-based investment project feasibility intelligent analysis and decision support system according to claim 1 is characterized in that: The node feature update formula of the graph neural network in the dynamic knowledge graph construction module is: in, For the The weight matrix of the layer, is the bias vector, is the activation function.

5. The AI-based investment project feasibility intelligent analysis and decision support system according to claim 1 is characterized in that: The industry volatility index calculation includes: calculating the financial volatility index and the market volatility index of the enterprise node set in the industry, and determining the industry volatility index through a weighted formula, where is the industry status weight of node i.

6. The AI-based investment project feasibility intelligent analysis and decision support system according to claim 1 is characterized in that: In the deep reinforcement learning decision module, the feasibility prediction network uses temporal difference learning to update the Q-value function. The update formula is: in, are DQN network parameters, is the target network parameter, is the learning rate, is the discount factor.

7. The AI-based investment project feasibility intelligent analysis and decision support system according to claim 1 is characterized in that: The deep reinforcement learning decision module is embedded in the Monte Carlo tree search, and the confidence upper bound formula is adopted in the node selection phase. In the simulation phase, the investment return distribution is generated based on the policy network, and the number of node visits and the average reward value are updated through backpropagation.

8. The AI-based investment project feasibility intelligent analysis and decision support system according to claim 1 is characterized in that: In the explainable AI module, the SHAP value calculation is solved by an approximate algorithm, where the baseline value is the predicted mean of the training set, and the visual display includes a SHAP value heat map of risk keywords and a contribution bar chart of structured features.

9. An AI-based investment project feasibility intelligent analysis and decision support method, comprising an AI-based investment project feasibility intelligent analysis and decision support system according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1. Data collection and preprocessing: Collect structured and unstructured data, and perform cleaning, standardization, word segmentation, and denoising preprocessing respectively; S2. Multimodal Data Fusion: Utilize the BERT variant model to extract unstructured data feature vectors, and then concatenate or weightedly fuse them with structured data feature vectors to generate multimodal fused data vectors. S3. Dynamic Knowledge Graph Construction: Based on multimodal fusion data, we use graph neural networks to build an industry upstream and downstream association graph, and update node features and connection weights based on market data fluctuations. S4. Investment Project Evaluation: Multimodal fusion data and dynamic knowledge graphs are input into the deep reinforcement learning decision-making module. Feasibility is predicted through the feasibility prediction network, and risk is assessed through the risk assessment network. An evaluation report is generated based on the Monte Carlo tree search simulation results. S5. Decision interpretation and compliance checking: Visualize key decision-making basis through the SHAP value attribution system and generate audit trail logs according to regulatory requirements; S6. Model update and optimization: Based on the deviation between actual investment results and predicted results, the model parameters are adjusted through the online learning mechanism, and the T+1 model update is triggered in conjunction with industry risk warnings.

10. The AI-based investment project feasibility intelligent analysis and decision support method according to claim 9, characterized in that: In step S2, the text sequence is subjected to layer normalization and multi-layer attention mechanism processing to generate a text feature vector, and a risk score is calculated based on the keyword position and context weight, and the score is quantified and converted into a feature vector.

Images