Pre-sale fund supervision method and device based on multi-objective optimization

Abstract

The invention discloses a presale fund supervision method and device based on multi-objective optimization, and relates to the field of financial risk management. Constructing a dynamic knowledge graph by fusing multi-source data, and forming an entity relationship network of a project, an enterprise and an associated party to quantify a risk path; predicting the comprehensive risk probability by adopting multiple models integrating static, time sequence and graph neural networks; the probability is input into an optimization decision engine, an improved multi-objective evolutionary algorithm is used to solve a Pareto optimal solution set of fund security, flow efficiency and project delivery rate, and objective weights are adjusted according to dynamic constraints; and finally, matching an optimal solution based on decision preference through a man-machine cooperation interface, automatically generating a personalized early warning and response strategy, and deploying and executing the personalized early warning and response strategy. According to the invention, dynamic and accurate supervision of presale funds is realized, and the fund security and the project delivery guarantee level are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of financial risk management, and in particular to a method, apparatus, equipment and storage medium for the supervision of pre-sale funds based on multi-objective optimization. Background Technology

[0002] The supervision of pre-sale funds for commercial housing is a core aspect of risk prevention and control in the real estate market, aiming to prevent developers from misappropriating funds and causing projects to fail. Traditional supervision methods rely on manual review and single-system monitoring, which suffers from inefficiency, insufficient accuracy, and lack of transparency. Currently, the pre-sale fund supervision system for commercial housing is evolving from static control to intelligent dynamic early warning, aiming to balance multiple objectives such as fund security, utilization efficiency, and project progress. While existing technical solutions have established a basic supervision framework, there is still room for improvement in the dynamic adaptability of early warning thresholds and the refinement of response strategies.

[0003] 1. The "Trust-based Prepaid Consumption Supervision System and Method" jointly developed by National Trust and others introduces a "payment + trust + bank" model, strengthening fund security through multi-level account technology. While it optimizes information flow and fund transfer, it primarily focuses on prepaid consumption scenarios and is not applicable to multi-objective optimization in the field of pre-sale fund supervision.

[0004] 2. Regulatory solutions based on multi-source data fusion: For example, the solution from China Power Construction Smart Cloud Data uses a dynamic semantic aligner to link bank statements, ERP events, and unstructured documents, and utilizes graph neural networks for anomaly detection. While this method achieves multi-source data integration, it focuses on general financial risk control and is not optimized for pre-sale fund supervision scenarios. Furthermore, it lacks key dimensions such as project progress verification, making it difficult to comprehensively cover the needs of pre-sale fund supervision. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] To address this, this invention proposes a pre-sale fund supervision method based on multi-objective optimization, belonging to the field of financial risk management. It quantifies risk transmission paths by constructing a dynamic knowledge graph; predicts the overall project risk probability using a multi-model fusion method; solves for the Pareto optimal solution set of fund security, liquidity efficiency, and project delivery rate using a multi-objective evolutionary algorithm; and generates personalized early warning and response strategies through human-machine collaborative decision-making, achieving automated dynamic supervision and effectively improving the level of fund security and project delivery assurance.

[0007] Another objective of this invention is to propose a pre-sale fund monitoring device based on multi-objective optimization.

[0008] The third objective of this invention is to provide a computer device.

[0009] The fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0010] To achieve the above objectives, this invention proposes a pre-sale fund supervision method based on multi-objective optimization, comprising: S1, construct a dynamic knowledge graph, and through the fusion and entity processing of multi-source heterogeneous data, form an entity relationship network including regulated projects, real estate development companies, regulatory accounts and related parties, and quantify the risk transmission path; S2, based on dynamic knowledge graphs and multi-source heterogeneous data, adopts a multi-model fusion prediction method that integrates static feature models, time series models and graph neural networks to generate comprehensive risk probability values ​​for regulatory projects; S3, input the comprehensive risk probability value into the multi-objective optimization decision engine, solve the Pareto optimal solution set of capital security, liquidity efficiency and project delivery success rate through the improved multi-objective evolutionary algorithm, and adjust the optimization objective weights according to dynamic constraints. S4. The Pareto optimal solution set is displayed through the human-machine collaborative decision-making interface. The optimal solution is matched based on the preference weights set by the decision-maker, and a personalized early warning threshold vector and response strategy combination are generated and deployed to the automated execution module to achieve dynamic monitoring.

[0011] The pre-sale fund supervision method based on multi-objective optimization according to an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, the construction of a dynamic knowledge graph, through multi-source heterogeneous data fusion and entityization processing, forms an entity relationship network including regulated projects, real estate development enterprises, regulatory accounts, and related parties, and quantifies risk transmission paths, including: S11, based on the preset regulatory entity definition, uses entity linking and relationship extraction technology to fuse multi-source heterogeneous data and generate a dynamic knowledge graph containing node attributes and edge relationships. S12 uses a graph computing engine to calculate node centrality, community clustering, and risk transmission paths in real time, and outputs graph feature vectors to a multi-objective optimization decision engine.

[0012] In one embodiment of the present invention, the step of generating a comprehensive risk probability value for a regulatory project based on dynamic knowledge graphs and multi-source heterogeneous data, using a multi-model fusion prediction method integrating static feature models, time-series models, and graph neural networks, includes: S21 employs an improved NSGA-III algorithm for multi-objective evolutionary optimization, using non-dominated sorting and a reference point-based selection mechanism to solve for the Pareto optimal solution set. S22, based on dynamic weight allocation formula The prediction results of the three sub-models are weighted and fused to generate a comprehensive risk probability value; among them, , The normalized weighted scores for the three sub-models are: This is a scaling factor used to control the degree of differentiation in weight allocation.

[0013] In one embodiment of the present invention, the step of inputting the comprehensive risk probability value into a multi-objective optimization decision engine, solving for the Pareto optimal solution set of capital security, liquidity efficiency, and project delivery success rate through an improved multi-objective evolutionary algorithm, and adjusting the optimization objective weights according to dynamic constraints includes: S31 uses Monte Carlo simulation to evaluate the system configuration schemes represented by each individual in the population and calculates the performance values ​​of the corresponding system configuration schemes on the three objective functions of capital security, liquidity efficiency and project delivery success rate. S32 adjusts the weights of optimization targets in real time based on preset dynamic constraints, including bottom-line constraints on capital security, constraints on strategy feasibility, and constraints on policy guidance.

[0014] In one embodiment of the present invention, the step of displaying the Pareto optimal solution set through a human-machine collaborative decision-making interface, matching the optimal solution based on the preference weights set by the decision-maker, generating a personalized early warning threshold vector and response strategy combination, and deploying it to the automated execution module to achieve dynamic monitoring includes: S41 uses a three-dimensional scatter plot to display the Pareto optimal solution set, where the color depth of the points is inversely proportional to the density of the local solution set, in order to assist decision-makers in making diverse choices. S42, by calculating weighted cosine similarity Quantify the degree of alignment between each solution and the decision-maker's subjective intent, and highlight and recommend the optimal compromise solution with the highest similarity; among which, This represents the preference weight vector set by decision-makers, reflecting the degree of importance attached to different objectives. Indicates the first A vector of objective function values ​​for each Pareto solution.

[0015] In one embodiment of the present invention, it further includes: S5, through the continuous learning and feedback optimization module, feeds back the actual effect data after the strategy is executed to the risk prediction module and the multi-objective optimization decision engine. Based on the incremental training method, the prediction model is fine-tuned, and the optimizer is updated through reinforcement learning using the multi-objective multi-armed gambling machine algorithm.

[0016] To achieve the above objectives, another aspect of the present invention proposes a pre-sale fund monitoring device based on multi-objective optimization, comprising: The dynamic knowledge graph construction module is used to build dynamic knowledge graphs. Through the fusion and entity processing of multi-source heterogeneous data, it forms an entity relationship network that includes regulated projects, real estate development companies, regulatory accounts and related parties, and quantifies the risk transmission path. The multi-model fusion prediction module is used to generate a comprehensive risk probability value for regulatory projects based on dynamic knowledge graphs and multi-source heterogeneous data, using a multi-model fusion prediction method that integrates static feature models, time series models and graph neural networks. The multi-objective optimization decision module is used to input the comprehensive risk probability value into the multi-objective optimization decision engine, solve the Pareto optimal solution set of capital security, liquidity efficiency and project delivery success rate through an improved multi-objective evolutionary algorithm, and adjust the optimization objective weights according to dynamic constraints. The human-machine collaboration and dynamic monitoring module is used to display the Pareto optimal solution set through the human-machine collaborative decision-making interface, match the optimal solution based on the preference weight set by the decision-maker, generate personalized early warning threshold vectors and response strategy combinations, and deploy them to the automated execution module to achieve dynamic monitoring.

[0017] In one embodiment of the present invention, it further includes: The continuous learning and feedback optimization module is used to feed back the actual effect data after the strategy is executed to the risk prediction module and the multi-objective optimization decision engine. It fine-tunes the parameters of the prediction model based on the incremental training method and updates the optimizer through reinforcement learning using the multi-objective multi-armed gambling machine algorithm.

[0018] The pre-sale fund supervision method and apparatus based on multi-objective optimization of this invention achieves precise quantification of risk transmission paths by constructing a dynamic knowledge graph, improves the comprehensiveness and accuracy of risk assessment by employing a multi-model fusion prediction method, and balances multiple objectives such as fund security, liquidity efficiency, and project delivery using a multi-objective optimization decision engine. It effectively overcomes the limitations of traditional supervision methods, such as lagging risk identification, single decision-making dimensions, and lack of dynamic adjustment capabilities. It achieves closed-loop automated supervision from risk perception and intelligent decision-making to strategy execution, significantly improving the systematicness, accuracy, and adaptability of pre-sale fund supervision, and providing reliable technical support for preventing financial risks and ensuring project delivery.

[0019] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing a pre-sale fund supervision method based on multi-objective optimization as described in the first aspect embodiment.

[0020] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a pre-sale fund supervision method based on multi-objective optimization as described in the first aspect embodiment.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a pre-sale fund supervision method based on multi-objective optimization according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a pre-sale fund supervision device based on multi-objective optimization according to an embodiment of the present invention; Figure 3 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following description, with reference to the accompanying drawings, describes a method, apparatus, device, and storage medium for monitoring pre-sale funds based on multi-objective optimization according to embodiments of the present invention.

[0026] The core idea of ​​this invention is to construct a dynamic knowledge graph that integrates multi-source heterogeneous data, mapping regulated projects, development companies, regulated accounts, and related parties into an entity relationship network, thereby achieving precise quantification of the transmission path of financial risks. Based on this graph and multi-source data, the system employs a multi-model fusion method integrating static features, time series analysis, and graph neural networks to generate a comprehensive risk probability value for each project. This probability value is then input into a multi-objective optimization decision engine, which uses an improved multi-objective evolutionary algorithm to solve for the Pareto optimal solution set of fund security, liquidity efficiency, and project delivery success rate, and adaptively adjusts the objective weights based on dynamic constraints. Finally, the solution set is displayed through a human-machine collaborative decision-making interface. Based on the decision-maker's preferences, the optimal solution is matched, and personalized early warning thresholds and response strategy combinations are automatically generated and deployed to the execution module to achieve closed-loop dynamic supervision. This transforms the traditional static supervision model into an adaptive control system that deeply perceives risks, intelligently balances objectives, and continuously optimizes, significantly improving the accuracy, balance, and system efficiency of fund supervision.

[0027] Example 1 To achieve the above invention, embodiments of the present invention provide a pre-sale fund supervision method based on multi-objective optimization, such as... Figure 1 As shown, it includes: S1 constructs a dynamic knowledge graph, which forms an entity relationship network including regulated projects, real estate development companies, regulatory accounts and related parties through the fusion and entityization of multi-source heterogeneous data, and quantifies the risk transmission path.

[0028] Specifically, this step integrates multimodal information from regulatory business systems, bank statements, IoT devices, enterprise-reported data, and publicly available external data to form an entity relationship network with "regulated projects" as the core entity. Specifically, the system first defines the ontology structure of the knowledge graph, clarifying node types (such as real estate development enterprises, regulatory accounts, and regulatory agencies) and edge types (such as holding companies, guarantees, fund transfers, and engineering contracting), thereby constructing a semantically structured graph pattern layer.

[0029] Furthermore, the system employs entity linking and relationship extraction techniques to map structured and unstructured data onto a unified graph semantic model. For example, it uses Natural Language Processing (NLP) technology to parse unstructured data such as contract texts and public opinion information, extracting entities and their relationships, and aligning them with entities such as projects, accounts, and enterprises in the structured data. After the graph is instantiated, the system uses stream processing technologies (such as Apache Flink or Kafka Streams) to access business change data in real time, driving the dynamic updates of the graph topology and entity attributes to ensure that the graph always reflects the latest regulatory status.

[0030] Specifically, the system quantifies risk transmission paths through a graph computing engine. For example, it calculates graph features such as node centrality, modularity, and risk propagation strength to identify key risk nodes and their potential impact range. The calculation of risk propagation strength can be based on the weights of relationships between nodes and path lengths, forming a quantitative assessment index of risk propagation.

[0031] Specifically, this dynamic knowledge graph is widely used in risk warning, regulatory strategy formulation, and systemic risk identification. For example, when the parent company of a real estate development company experiences a cash flow risk, the system can quickly identify its related projects through the graph and assess the strength and speed of the risk transmission path, thereby triggering corresponding early warning mechanisms and response strategies.

[0032] Specifically, by constructing a knowledge graph with dynamic evolution capabilities, the system can achieve in-depth correlation analysis of regulated objects and their related parties, significantly improving the breadth and accuracy of risk identification, providing structured and semantic risk transmission basis for subsequent multi-objective optimization decisions, and enhancing the intelligence level and responsiveness of the regulatory system.

[0033] Furthermore, S1 includes: S11, based on the preset definition of regulatory entities, uses entity linking and relationship extraction technology to fuse multi-source heterogeneous data and generate a dynamic knowledge graph containing node attributes and edge relationships.

[0034] Specifically, the technical implementation principle of this step is based on ontology modeling, entity recognition and relation extraction technologies in knowledge graph construction, combined with the dynamic update mechanism of graph database, to achieve structured modeling and real-time evolution of regulatory objects and their related entities.

[0035] Specifically, the system first defines an ontology structure centered on "regulated projects," expanding to include entity categories such as real estate development companies, regulated accounts, construction companies, collateral, homebuyers, and regulatory agencies. It also pre-defines relationship types between entities, such as holding companies, guarantees, fund transfers, and engineering contracting, forming a semantic pattern layer of the graph. Subsequently, the system uses entity linking technology to uniformly map and align entity instances from structured business systems (such as bank statements and ERP systems) with unstructured data sources (such as contract texts and public opinion information). For example, it uses a Named Entity Recognition (NER) model to identify mentions of "a certain developer" in text and links them to a unique identifier (such as a unified social credit code), ensuring consistency of entities across data sources.

[0036] Furthermore, relation extraction is based on a combination of rules and deep learning to extract semantic relationships between entities from text. For example, it identifies "Party A and Party B signed an engineering contract" from a contract and transforms it into the edge relationship "Engineering Contract (Party A → Party B)" in a graph. During relation modeling, the system supports multiple relation types and adds attribute information to each edge, such as relation strength, timestamp, and amount, to enhance the expressive power of the graph.

[0037] Furthermore, the system supports a dynamic update mechanism, using a stream processing engine (such as Apache Flink) to access business change data in real time, driving the evolution of the graph topology and entity attributes. The graph update frequency can be set according to the characteristics of the data source; for example, bank transaction data can be set to update hourly, and public opinion data can be set to update every 15 minutes. In addition, the system has a built-in graph computing engine that supports computing node centrality (such as PageRank), community clustering (such as the Louvain algorithm), and risk transmission paths (such as shortest path analysis). Its output results serve as input features for a multi-objective optimization engine, used for the quantitative assessment of risk transmission intensity.

[0038] Specifically, this dynamic knowledge graph is widely used in risk transmission analysis, related party risk identification, and regulatory strategy matching. For example, when a developer's parent company experiences a cash flow risk, the system can quickly identify its related projects through the graph and assess potential risk transmission paths, thereby triggering corresponding early warning and response strategies.

[0039] Specifically, by constructing a knowledge graph with semantic association and dynamic evolution, the system can realize multi-dimensional profiling of regulatory objects and visualized analysis of risk transmission paths, providing high-quality and highly interpretable graph structure feature inputs for subsequent multi-objective optimization, and significantly improving the intelligence level and risk identification capability of the regulatory system.

[0040] S12 uses a graph computing engine to calculate node centrality, community clustering, and risk transmission paths in real time, and outputs graph feature vectors to a multi-objective optimization decision engine.

[0041] Specifically, the graph computing engine performs topological analysis based on a dynamic knowledge graph constructed from graph databases (such as Neo4j and JanusGraph). Nodes include, but are not limited to, "regulated projects," "real estate development companies," "regulated accounts," "construction companies," and "homebuyers," while edges represent semantic connections such as "fund transfers," "contractual relationships," and "guarantee relationships." The graph computing engine employs a distributed graph processing framework (such as Apache Flink Gelly and GraphX) to support streaming data access and real-time updates to the graph structure. During the computation process, the system first evaluates the centrality of nodes using algorithms such as PageRank, Katz Centrality, or Eigenvector Centrality to identify entities with high influence or key transmission roles in the graph. Second, it performs community clustering analysis using the Louvain algorithm or Label Propagation algorithm to identify closely related entity groups, thereby revealing potential areas of systemic risk aggregation. Finally, the system identifies the transmission paths of risk in the graph using shortest path algorithms (such as Dijkstra's algorithm and BFS) or risk transmission strength models (such as propagation mechanisms based on graph attention networks) to assess the likelihood and intensity of risk event diffusion among related entities.

[0042] Furthermore, in centrality calculations, the damping coefficient of PageRank is typically set to 0.85, and the attenuation factor β of KatzCentrality is generally between 0.01 and 0.1 to avoid excessive accumulation of path weights. In community clustering algorithms, modularity, as a core indicator for evaluating cluster quality, should have a value greater than 0.3, indicating statistical significance. In risk transmission path analysis, the path length threshold is typically set to no more than 3 hops (i.e., 3-degree association) to ensure computational efficiency and risk relevance.

[0043] Specifically, this step is widely used in real estate pre-sale fund supervision systems, especially in complex supervision networks involving multiple projects, multiple enterprises, and multiple regulatory bodies. For example, when a developer's parent company experiences a cash flow risk, the graph computing engine can quickly identify the risk transmission path in its subsidiary projects and output graph feature vectors to a multi-objective optimization engine, thereby dynamically adjusting the early warning thresholds and response strategies for relevant projects to achieve early intervention and precise handling of risks.

[0044] Specifically, this step, through the graph feature vectors extracted by the graph computing engine, can effectively reveal the nonlinear transmission mechanism of risk and the implicit correlations between entities, providing input variables with high interpretability and predictive value for the multi-objective optimization engine. Its technical value lies in improving the system's ability to identify systemic risks, enhancing the foresight and adaptability of regulatory strategies, thereby realizing a shift from a "passive response" to a "proactive prediction" regulatory model.

[0045] S2, based on dynamic knowledge graphs and multi-source heterogeneous data, adopts a multi-model fusion prediction method that integrates static feature models, time series models and graph neural networks to generate comprehensive risk probability values ​​for regulatory projects.

[0046] Specifically, this step constructs a comprehensive risk assessment system with high robustness and adaptability by integrating the prediction results of multiple models.

[0047] Specifically, this step first relies on the dynamic knowledge graph constructed in step S1 and the fusion results of multi-source heterogeneous data. Static feature models (such as XGBoost or LightGBM) process cross-sectional data such as developer qualifications and project types, improving the model's ability to identify baseline risks through optimal binning and feature importance weighted sampling mechanisms. Time series models (such as LSTM or TCN) focus on time-series data such as cash flow and project progress, using dilated causal convolution and gating mechanisms to model long-term dependencies, thereby capturing the dynamic evolution trend of risks. Graph neural networks (GNNs) are based on the node and edge relationships in the knowledge graph, quantifying the asymmetric transmission effect of risk between related entities through a hierarchical graph attention propagation mechanism.

[0048] Furthermore, the prediction results from each sub-model are fused using a dynamic weighting mechanism. (Weights) The calculation depends on the model performance score. Its formula is: ; in, From prediction accuracy Data timeliness and result stability Joint decision, This is a scaling factor used to adjust the sensitivity of weight allocation. The final overall risk probability value. Output from each sub-model The weighted sum is given as follows: .

[0049] Specifically, this step is widely used in real estate pre-sale fund supervision systems, particularly in scenarios such as project risk level assessment, fund disbursement approval decisions, and regulatory strategy matching. The system dynamically updates its knowledge graph and feature library by real-time access to bank statements, project progress, and public opinion data, ensuring that the predictive model is always based on the latest data for risk assessment.

[0050] Specifically, through a multi-model fusion mechanism, the system can integrate static, time-series, and graphical information to achieve multi-dimensional, dynamic, and accurate quantification of risks in regulated projects. Compared to a single model, this method significantly improves the stability and adaptability of predictions, reduces false positives and false negatives, and provides reliable risk input for subsequent multi-objective optimization decisions, serving as a key support for achieving "proactive supervision" and "intelligent intervention."

[0051] Furthermore, S2 includes: S21 employs an improved NSGA-III algorithm for multi-objective evolutionary optimization, using non-dominated sorting and a reference point-based selection mechanism to solve for the Pareto optimal solution set.

[0052] Specifically, the algorithm is used to optimize the combination of early warning thresholds and response strategies in the pre-sale fund supervision system to achieve synergistic optimization of fund security, utilization efficiency and project delivery success rate.

[0053] Furthermore, NSGA-III first generates a population of N individuals through initialization. Each individual represents a complete system configuration scheme, including personalized early warning threshold vectors for each item. and policy mapping matrix Subsequently, the system performs simulation evaluations on each individual based on a digital twin environment or historical data, calculating its performance on three objective functions. (Risk minimization) (Maximizing the efficiency of capital utilization) and The performance value is maximized on the delivery probability. During the evolutionary iteration process, the algorithm performs selection, crossover, and mutation operations to generate offspring populations. NSGA-III divides the population into multiple levels through non-dominated sorting and combines it with crowding entropy calculation based on reference points to ensure that the solution set has good distribution and diversity on the Pareto front.

[0054] Furthermore, key parameters of the algorithm include population size. Crossover probability Probability of mutation and the number of reference points The reference point is selected based on the priority distribution of regulatory objectives, and is usually set as follows: This is to cover a reasonable distribution of the multidimensional target space. Crossover and mutation operations use real-number encoding, and the mutation operator can optionally use Gaussian mutation, with its standard deviation... Typically set to To balance exploration and development capabilities.

[0055] Specifically, the algorithm is deployed in a multi-objective optimization decision engine to generate Pareto optimal solutions, which are then used by regulatory decision-makers for human-machine collaborative selection via a visual interface. Through this mechanism, the system can dynamically adjust regulatory strategies based on real-time market conditions, policy guidance, and project status, achieving a shift from "passive alerting" to "proactive intervention."

[0056] Specifically, through the non-dominated sorting and reference point selection mechanism of NSGA-III, the system can find the optimal compromise between multiple conflicting objectives, thereby improving the scientific nature, flexibility and foresight of supervision, and providing solid algorithmic support for achieving refined and intelligent pre-sale fund supervision.

[0057] S22, based on dynamic weight allocation formula The prediction results of the three sub-models are weighted and fused to generate a comprehensive risk probability value; among them, , The normalized weighted scores for the three sub-models are: This is a scaling factor used to control the degree of differentiation in weight allocation.

[0058] Specifically, the core technical principle of this step is to introduce an adjustable weight allocation mechanism, which enables the system to dynamically adjust the contribution weight of each sub-model in the final risk assessment based on its predictive performance, data timeliness, and result stability within a specific time window, thereby improving the accuracy and robustness of comprehensive risk prediction.

[0059] Specifically, the system first evaluates the performance of static risk prediction sub-models (such as XGBoost and LightGBM), dynamic risk prediction sub-models (such as LSTM or TCN), and graph risk propagation sub-models (such as GNN). Evaluation metrics include prediction accuracy. Data timeliness and the variance of the prediction results And according to the preset weighting coefficients Calculate the normalized model score. Then, the score of each sub-model is calculated using the formula described above. Mapped to a dynamic weight between 0 and 1 ,in This is a scaling factor used to control the degree of differentiation in weight allocation. When When the value is larger, the model with the higher score will receive more significant weight, thus dominating the fusion process.

[0060] Furthermore, It is usually set to a positive real number, with a recommended range of [0.5, 2.0]. Its value can be adjusted according to the system's tolerance for model stability. The calculation cycle can be set daily or every 7 days to ensure that the dynamic updates of the model weights are consistent with data freshness. Furthermore, the exponential function in this formula... The non-negativity and normalization properties of the weights are guaranteed, which gives the fusion result good mathematical properties and interpretability.

[0061] Specifically, in the pre-sale fund supervision system, three sub-models predict project risks from the perspectives of static characteristics, temporal behavior, and the transmission of associated risks, respectively, and their output results... Typically, the probability value is between 0 and 1. Through dynamic weight fusion, the system can adaptively enhance the influence of the better-performing model, thereby generating a more representative and robust comprehensive risk probability value. This is used for subsequent multi-objective optimization decisions.

[0062] Specifically, its weight allocation mechanism not only relies on the model's prediction accuracy but also comprehensively considers data timeliness and prediction stability, thereby avoiding the model obsolescence or prediction drift problems that may occur in traditional fixed-weight fusion methods. This step not only improves the overall prediction performance of the system but also enhances its adaptability in complex and ever-changing regulatory environments, serving as a crucial support for achieving intelligent, dynamic, and multi-objective regulatory decision-making.

[0063] S3. Input the comprehensive risk probability value into the multi-objective optimization decision engine, solve the Pareto optimal solution set of capital security, liquidity efficiency and project delivery success rate through the improved multi-objective evolutionary algorithm, and adjust the optimization objective weights according to dynamic constraints.

[0064] Specifically, this step is based on an improved multi-objective evolutionary algorithm (MOEA), which aims to find the Pareto optimal solution set among capital security, liquidity efficiency and project delivery success rate, and adjust the weights of the optimization objectives according to dynamic constraints, thereby achieving personalized and dynamic adaptation of regulatory strategies.

[0065] Specifically, this step first will As input variables, and combining basic project attributes (such as project size, developer credit rating, and construction schedule) with external environmental variables (such as market fluctuations and policy changes), a decision space for a multi-objective optimization problem is constructed. The objective function of the optimization problem includes minimizing the total system risk. Maximize the efficiency of fund utilization And to maximize the probability of on-time project delivery. .in, As project weight, To improve capital utilization, This represents the delivery probability.

[0066] Furthermore, the system employs an improved NSGA-III algorithm. This algorithm, through non-dominated sorting and a diversity maintenance mechanism based on reference points, can maintain the distribution and convergence of the solution set in a high-dimensional objective space. During optimization, each "individual" represents a combination of regulatory thresholds and response strategies. The algorithm evaluates its performance on the three objectives by simulating future capital flows and project progress. During iteration, the system introduces a dynamic weight adjustment mechanism, automatically adjusting the weights of each dimension in the objective function based on real-time regulatory policies, changes in the market environment, and project risk status to reflect the current regulatory priority.

[0067] Specifically, the population size of NSGA-III The number of iterations is usually set to 100-200. The mutation probability is 50-100. The crossover probability is between 0.1 and 0.2. The value is 0.8~0.9. Number of reference points. The value is typically set to 5 to 10 to ensure a uniform distribution of the solution set in the target space.

[0068] Specifically, it can be deployed on city-level or regional-level pre-sale funds supervision platforms for commercial housing, supporting parallel optimization of multiple projects under construction. By outputting Pareto optimal solutions, supervisors can select the strategy combination that best meets the current supervisory objectives through a visual interface, realizing a shift from "passive response" to "proactive intervention." Its technological value lies in the fact that, through the combination of mathematical modeling and intelligent optimization algorithms, the system can generate optimal supervisory strategies for each project under complex constraints, improving the scientific nature, flexibility, and response efficiency of supervision.

[0069] Furthermore, S3 includes: S31 uses Monte Carlo simulation to evaluate the system configuration schemes represented by each individual in the population, and calculates the performance values ​​of the corresponding system configuration schemes on three objective functions: capital security, liquidity efficiency, and project delivery success rate.

[0070] Specifically, this step, based on the evolutionary computation framework of the NSGA-III algorithm, quantitatively evaluates the performance of each candidate solution (i.e., individual) on three objective functions: capital security, liquidity efficiency, and project delivery success rate, thereby providing a basis for subsequent non-dominated ranking and selection operations.

[0071] Furthermore, Monte Carlo simulations, by constructing digital twin environments or based on historical datasets, randomly sample and extrapolate regulatory configuration schemes represented by each individual. The simulation process typically includes random disturbances in fund flow paths, modeling uncertainties in project progress, and random changes in the external market environment. The system employs a time-series simulation method, dividing a future planning cycle (e.g., a quarter) into several time steps (e.g., weeks). In each time step, based on the threshold vector set for the current individual... and policy mapping matrix The simulation aims to determine the probability of fund disbursement, project progress, and risk events. During the simulation, random variables are introduced to reflect uncertainties in real-world business operations, such as the volatility of fund inflows and the probability of construction delays.

[0072] Specifically, the performance evaluation of the Monte Carlo simulation depends on the calculation results of three objective functions. The objective of financial security is... Through weighted risk probability An assessment was conducted, in which As project weight, Let $\frac{i}{i}$ represent the probability of a risk event occurring in project $i$ during the simulation. Liquidity efficiency objective. The fund utilization rate is used as a metric to reflect the efficiency of fund use. Project delivery success rate target. Then the output of the prediction model An assessment is conducted to represent the probability of the project being delivered on time within the simulation period. The number of simulations is typically set to [number missing]. This is done twice to ensure the stability of the statistical results.

[0073] Specifically, this step is widely applied in the optimization decision-making process of the real estate pre-sale fund supervision system. For example, when regulatory agencies formulate quarterly regulatory strategies, the system will conduct Monte Carlo simulations on multiple candidate schemes to assess their risk control capabilities and fund utilization efficiency under different market environments. The simulation results can provide regulators with intuitive decision support, helping them find the optimal balance between fund security and utilization efficiency.

[0074] Specifically, through high-fidelity simulation evaluation, the system can accurately quantify the performance of each individual in the multi-objective space, thus providing a reliable fitness value for the NSGA-III algorithm. This not only improves the convergence speed and solution distribution of the optimization process, but also enhances the system's adaptability to complex regulatory scenarios, providing a solid foundation for achieving dynamic, intelligent, and personalized pre-sale fund supervision.

[0075] S32 adjusts the weights of optimization targets in real time based on preset dynamic constraints, including bottom-line constraints on capital security, constraints on strategy feasibility, and constraints on policy guidance.

[0076] Specifically, the technical implementation of this step is based on an adaptive adjustment strategy of dynamic constraint perception and objective function weights to achieve the optimal response of the regulatory system in complex and ever-changing environments.

[0077] From a technical implementation perspective, this step introduces a constraint-aware mechanism to map key constraints such as external regulatory policies, system resource limitations, and project funding safety limits into a multi-objective optimization model in real time. The system first defines three types of constraints: funding safety limit constraints... Constraints on the feasibility of strategies and policy guidance constraints .in, As a hard constraint, ensure that the balance of the regulated account is not lower than the statutory minimum; Limit the implementation costs and resource consumption of the strategy, such as approval levels and frequency of manual intervention; These are soft constraints, reflecting the phased priorities of regulatory policies, such as the increased priority given to delivery success rates under the "guaranteed delivery" policy. The system uses a constraint detection module to evaluate in real time the degree to which the current environment meets various constraints, forming a constraint state vector. ,in Representing constraints Satisfaction level.

[0078] Specifically, the system is based on the constraint state vector Dynamically adjust the weight vector of the objective function ,in , , These correspond to the objectives of minimizing risk, maximizing capital utilization efficiency, and maximizing delivery probability, respectively. The weight adjustment mechanism combines linear interpolation with exponential decay to ensure that the priority of the corresponding objective is rapidly increased when constraints exceed limits. For example, when... ( When the preset safety threshold is used, It will be linearly amplified, while and The amount will be reduced proportionally to prioritize the safety of funds.

[0079] Specifically, this step is widely used in real estate pre-sale fund supervision systems, especially in scenarios such as sudden changes in project risk, policy adjustments, or market fluctuations. For example, when a developer's negative public opinion leads to increased risk transmission to its related projects, the system will automatically upgrade... The strength of the constraints allows for prioritizing delivery assurance strategies during optimization. Furthermore, when regulatory resources are strained, Limiting the complexity of the strategy will encourage the system to choose an optimized solution that requires minimal intervention and is highly efficient.

[0080] Specifically, this step significantly improves the adaptability and robustness of the multi-objective optimization model by introducing a dynamic constraint perception mechanism. Its core value lies in enabling flexible adjustments to regulatory objectives, allowing the system to automatically find the optimal balance point under different constraints, thereby improving overall regulatory effectiveness. This mechanism not only enhances the system's real-time response capability but also provides regulators with a more operational strategy selection space, serving as a key technological support for achieving the dual objectives of "intelligent regulation" and "risk controllability."

[0081] S4. The Pareto optimal solution set is displayed through the human-machine collaborative decision-making interface. The optimal solution is matched based on the preference weights set by the decision-maker, and a personalized early warning threshold vector and response strategy combination are generated and deployed to the automated execution module to achieve dynamic monitoring.

[0082] Specifically, the technical implementation principle of this step is based on multi-objective optimization theory and human-machine collaboration mechanism, aiming to realize the personalized generation and automated deployment of regulatory strategies.

[0083] Specifically, the system first presents the Pareto optimal solution set output by the multi-objective optimization engine to regulatory decision-makers through a visualization interface. This solution set consists of multiple non-dominated solutions, each representing a set of system configuration schemes that achieve an optimal balance among the three objectives of risk control, capital utilization efficiency, and project delivery probability. The system uses two visualization methods: a 3D scatter plot and a parallel coordinate axis, to display the objective function values ​​respectively. , , The system maps solutions onto coordinate axes, enabling decision-makers to intuitively understand the trade-offs between different solutions. To enhance interactivity, the system introduces a solution set density coloring mechanism, where the color intensity of a point is inversely proportional to the local solution set density, thereby highlighting solutions in sparse regions and facilitating decision-makers in identifying solutions with unique advantages.

[0084] Furthermore, the system supports decision-makers inputting preference weight vectors. ,in Let represent the preference strength for the i-th objective, and satisfy . The system calculates weighted cosine similarity. Each solution is matched with the decision-maker's preferences, and the solution with the highest similarity is recommended as the final execution plan. The threshold vector corresponding to this solution... With policy mapping matrix It will be extracted and deployed to the automated execution module.

[0085] Specifically, this step applies to dynamic strategy formulation scenarios within the real estate pre-sale fund supervision system. Supervisors can adjust weights in a visual interface based on current policy directions (such as "ensuring delivery"), market conditions (such as periods of tight funding), or project characteristics (such as high-risk developers). The system provides real-time feedback on the optimal solution, assisting them in making informed decisions. This mechanism is particularly suitable for scenarios involving large-scale, multi-project parallel supervision, significantly improving supervisory efficiency and response speed.

[0086] Specifically, this step achieves efficient transformation from multi-objective optimization solutions to actual regulatory strategies. Through a human-machine collaboration mechanism, the system combines quantitative analysis with qualitative judgment to ensure that the output warning thresholds and response strategies not only conform to mathematical optimality but also align with the actual management intentions of regulators. The resulting personalized threshold vectors and strategy combinations, once deployed to the automated execution module, enable real-time monitoring of regulatory accounts and strategy triggering, thereby forming a closed-loop regulatory system and enhancing the overall intelligence and adaptability of the system.

[0087] Furthermore, S4 includes: S41 uses a three-dimensional scatter plot to display the Pareto optimal solution set, where the color depth of the points is inversely proportional to the density of the local solution set, in order to assist decision-makers in making diverse choices.

[0088] Specifically, the system introduces a solution set density coloring mechanism to enhance visualization and assist decision-makers in making diverse choices. Specifically, the color depth of a point is inversely proportional to the local solution set density, i.e. This density can be calculated using kernel density estimation (KDE) or a local counting method based on a spatial grid. In practical implementations, the system can employ a K-nearest neighbor algorithm based on Euclidean distance for each point. Statistical analysis of the K nearest neighbors is performed to calculate the local density. This maps color values ​​accordingly. The color mapping function can be set as follows: ; in, This is a smoothing factor used to prevent division by zero errors; it is typically set to a value of [value missing]. to between.

[0089] Specifically, this 3D scatter plot is integrated into a human-machine collaborative decision-making interface, allowing regulators to select strategies under different market cycles or policy orientations. For example, when the real estate market experiences increased volatility, regulators can prioritize [specific strategies / targets]. In the (risk control) dimension, the system uses a color inverse mechanism to highlight solutions that perform well in this dimension but are not excessively sacrificed in other dimensions, thereby improving decision-making efficiency and scientific rigor.

[0090] Furthermore, by presenting the Pareto optimal solution set in a three-dimensional visualization and combining it with a density coloring mechanism, the distribution characteristics and diversity of the solution set are effectively revealed, preventing decision-makers from falling into local optima traps. Simultaneously, this method supports interactive exploration, incorporating weight adjustment and similarity matching formulas: ; The system can recommend the solution that best matches the decision-maker's subjective intention in real time, realizing a closed loop of intelligent decision-making through human-machine collaboration.

[0091] S42, by calculating weighted cosine similarity Quantify the degree of alignment between each solution and the decision-maker's subjective intent, and highlight and recommend the optimal compromise solution with the highest similarity; among which, This represents the preference weight vector set by decision-makers, reflecting the degree of importance attached to different objectives. Indicates the first A vector of objective function values ​​for each Pareto solution.

[0092] Specifically, this step is a key link in achieving deep integration of multi-objective optimization and human-machine collaborative decision-making, and its technical implementation principle is based on vector space model and preference matching mechanism.

[0093] Specifically, the system first maps each Pareto optimal solution to a standardized objective function vector. ,in , , These represent the normalized values ​​of risk control, liquidity, and regulatory costs (typically mapped to the [0,1] interval), respectively. Simultaneously, the system receives the preference weight vector set by the decision-maker in the visualization interface. This vector reflects the relative importance that decision-makers place on each regulatory objective. The weight vector must satisfy... This is to ensure that it is a legitimate preference distribution.

[0094] Furthermore, weighted cosine similarity The calculation relies on the dot product of the two vectors and the product of their magnitudes. Dot product This represents the degree of matching between the decision preference and the solution objective value in terms of direction, while the denominator is used for normalization to ensure that the similarity value is within the range of [-1, 1]. In practical systems, to avoid the case where the magnitude of the weight vector is zero, it is usually... Apply nonzero constraints, i.e. ,and Furthermore, the objective function vector The standardization process must follow a unified normalization standard, such as min-max scaling or Z-score standardization, to eliminate dimensional differences.

[0095] Specifically, this step is primarily used by regulatory decision-makers to choose options within a multi-objective trade-off framework. For example, in the context of tightening real estate market control policies, decision-makers may place greater emphasis on risk control. They will be given higher weight and the system will automatically highlight them for recommendation. The optimal solution is the one that best aligns with the decision-maker's intentions in terms of risk control. This mechanism supports real-time interaction; when the decision-maker adjusts the weights, the system immediately recalculates the similarity and updates the recommendation results, thus achieving dynamic decision support.

[0096] Specifically, this step transforms subjective decision-making into a calculable and traceable mathematical process by quantifying the matching degree of human-machine preferences, significantly improving the scientific nature and consistency of decision-making. Simultaneously, by automatically recommending optimal compromise solutions, the system reduces the cognitive burden on decision-makers and improves the efficiency and accuracy of regulatory responses, providing crucial technical support for achieving a closed-loop intelligent regulatory system.

[0097] S5, through the continuous learning and feedback optimization module, feeds back the actual effect data after the strategy is executed to the risk prediction module and the multi-objective optimization decision engine. Based on the incremental training method, the prediction model is fine-tuned, and the optimizer is updated through reinforcement learning using the multi-objective multi-armed gambling machine algorithm.

[0098] Specifically, the technical implementation principle of this step is based on the combination of incremental learning and reinforcement learning, aiming to continuously optimize the decision-making capabilities of the risk prediction model and the multi-objective optimization engine through the actual effect data after policy execution.

[0099] Furthermore, this module first collects actual performance data after strategy execution through the system's built-in data tracking mechanism, including key indicators such as whether risk events occurred, changes in fund utilization efficiency, and whether project progress improved. This data is stored in a structured manner and compared with the prediction results before strategy execution to form a feedback signal. This feedback signal serves as the reward vector for reinforcement learning, used to update the policy mapping matrix in the multi-objective optimization engine. and threshold vector This will improve the system's decision-making quality in future scenarios.

[0100] Specifically, the system employs a Multi-Objective Multi-Armed Bandit (MOMAB) algorithm to update the optimizer through reinforcement learning. This algorithm selects the optimal (threshold, strategy) combination in the action space through an explore-exploit mechanism. The reward function is defined as a multi-objective reward vector, including risk reduction rate, capital utilization improvement, and delivery success rate, and its scalarized form is:

[0101] in, Indicates the state Next action The obtained first The reward value for each objective. These are adaptive weights used to reflect the current regulatory environment's preference for each objective. The system periodically adjusts them based on actual performance. This enables dynamic responses to optimization objectives.

[0102] Specifically, this module is widely used in the closed-loop management process of real estate pre-sale fund supervision. For example, when a project is triggered by an abnormal outflow of funds and an alert is issued and a "suspension of non-construction fund payments" strategy is implemented, the system will track whether the strategy has effectively controlled the risk and whether it has affected the normal construction progress of the project, and use this feedback data for subsequent optimization engine strategy selection and threshold adjustment.

[0103] Specifically, by fine-tuning the parameters of the risk prediction model through incremental training, the system can continuously adapt to changes in the market environment and project status, avoiding model aging. At the same time, through reinforcement learning mechanisms, the optimization engine can gradually learn more effective strategy combinations, improving the accuracy and response efficiency of supervision, thereby building an intelligent supervision system with self-evolution capabilities.

[0104] This invention discloses a pre-sale fund supervision method based on multi-objective optimization. It achieves precise quantification of risk transmission paths by constructing a dynamic knowledge graph, improves the comprehensiveness and accuracy of risk assessment through multi-model fusion prediction, and balances multiple objectives such as fund security, liquidity efficiency, and project delivery using a multi-objective optimization decision engine. This method effectively overcomes the limitations of traditional supervision methods, such as lagging risk identification, single decision-making dimensions, and lack of dynamic adjustment capabilities. It achieves closed-loop automated supervision from risk perception and intelligent decision-making to strategy execution, significantly improving the systematicness, accuracy, and adaptability of pre-sale fund supervision, and providing reliable technical support for preventing financial risks and ensuring project delivery.

[0105] Example 2 To achieve the above invention, embodiments of the present invention also provide a pre-sale fund monitoring hardware based on multi-objective optimization, comprising: The hardware as a whole consists of five core modules that work together to form a complete intelligent closed loop from perception and decision-making to execution and feedback. The specific modules include: In one embodiment of the present invention, the multi-source heterogeneous data fusion and dynamic knowledge graph construction module includes: Specifically, this module is the foundation of the system's perception layer, responsible for the comprehensive collection, management, and deep correlation of data.

[0106] Specifically, the data acquisition layer includes internal business data acquisition and external environment data acquisition: Furthermore, internal business data: Through API interfaces, we connect with various business systems to obtain basic project information (scale, business type, location), detailed cash flow (receipts, disbursements, transfers), project construction progress plans and actual completions (obtained through IoT devices and digital supervision reports), sales filing data (number of units, area, amount), contract information, etc. in real time or near real time.

[0107] Furthermore, external environment data includes: enterprise profile data: obtaining developers' financial statements, credit ratings, equity structures, litigation information, administrative penalty records, and data on bond interest rates and stock fluctuations in the capital market from public channels; market and public opinion data: collecting macroeconomic indicators, regional real estate market price and volume data, policy and regulatory changes, and online news and social media public opinion (conducting sentiment analysis) about developers and projects through web crawling technology; and supply chain data: obtaining the operating conditions and public opinion information of major contractors and material suppliers.

[0108] Specifically, the data governance and feature engineering layer involves cleaning, denoising, standardizing, and normalizing the collected raw data. Based on this, a large feature library is built, including: static features (such as developer qualification levels), dynamic time-series features (such as the 7-day moving average of net capital inflows and project progress deviation rates), and derived risk indicators (such as the quick ratio and sales turnover cycle). The feature library construction process is as follows: First, the system performs "fusion and entityification of diverse heterogeneous data." It accesses and integrates multimodal data from regulatory business systems, bank statements, IoT sensors, enterprise-reported data, and publicly available external data through data interfaces. Based on pre-defined regulatory entities (such as "regulatory accounts," "development projects," and "real estate development enterprises"), data is associated and aligned to form a unified entity profile data foundation. Second, the system implements automated generation of multi-level feature engineering. It has a built-in feature operator library that automatically performs feature calculations based on the data foundation. This process generates features at three levels: Original features: Atomic indicators extracted directly from raw data; Derived features: Statistical, trend, and correlation features generated through techniques such as time series analysis and graph computation; Model features: Deep features output by pre-trained models.

[0109] Finally, the resume features a feature management system with a quality feedback loop. All generated features are stored in the feature library along with metadata (including data source, calculation logic, and update frequency), and are subject to real-time verification by the data monitoring module. The system innovatively introduces a "feature utility evaluation mechanism," which dynamically adjusts the feature calculation strategy by analyzing the contribution of features to multi-objective optimization decisions, forming a closed loop of "feature generation - quality evaluation - utility feedback - strategy optimization." This enables continuous self-optimization and refinement of the feature library, providing high-quality, highly interpretable dynamic feature input for the upper-level multi-objective optimization engine.

[0110] Specifically, the knowledge graph construction layer utilizes graph database technology to build a dynamic knowledge graph with "regulated projects" as the core entity. Nodes in the graph include: project companies, parent companies, contractors, suppliers, supervising banks, homebuyers, and regulatory agencies; edges represent the complex relationships between them, such as equity control, fund transfers, contractual relationships, and guarantee relationships. This graph can intuitively reveal risk transmission paths. For example, when other projects of a parent company encounter risks, the graph's associations can be used to quickly locate and assess their potential impact on the current regulated project. The construction process of the dynamic knowledge graph is as follows: Furthermore, the ontology layer is designed first. "Supervised Projects" is defined as the core entity, and related entities such as "Real Estate Development Enterprises," "Supervisory Accounts," "Construction Enterprises," "Collectibles," "Homebuyers," and "Supervisory Agencies" are expanded to include other entities. Simultaneously, relationship types between entities such as "Holding," "Guarantee," "Fund Transactions," and "Engineering Contracting" are pre-defined to form the semantic schema layer of the graph. Furthermore, the system performs multi-source data fusion and graph instantiation. From structured business data and unstructured text (such as public opinion and contracts), the system extracts entity instances and relational facts using entity linking and relationship extraction techniques, maps them to the ontology layer, and instantiates them in the graph database to generate an initial knowledge graph.

[0111] Specifically, the system ultimately achieves dynamic updates and risk transmission analysis of the graph. It utilizes stream processing technology to access real-time business change data, driving the dynamic evolution of the graph's topology and entity attributes. Innovatively, the system incorporates a graph computing engine that, based on this dynamic graph, calculates the intensity of risk transmission between entities in real time, identifies potential associated risks, and outputs the generated graph features (such as node centrality, community clustering, and risk transmission paths) to the multi-objective optimization engine. This provides in-depth insights into associated risks for intelligent optimization of thresholds and strategies.

[0112] In one embodiment of the present invention, the risk dynamic prediction module based on ensemble learning and time series models includes: This module utilizes the high-quality features and knowledge graphs output by the modules above to quantitatively predict the multidimensional risks of a project.

[0113] Specifically, risk definition and labeling: We clearly define risk events as: "the probability that within the next 90 days, the project will experience severe delays in progress (>30%) or a break in the funding chain (inability to pay project funds)." Based on historical data, each project is labeled with a risk tag (0 or 1) at each historical point in time.

[0114] Specifically, the multi-model fusion prediction engine includes: a static risk prediction sub-model: which adopts a gradient boosting decision tree (such as XGBoost, LightGBM) model, mainly processing static and cross-sectional features, such as developer qualifications and project type, to assess the baseline risk level of the project.

[0115] Furthermore, the data processing procedure for the static risk prediction sub-model is as follows: The data processing innovatively introduces an adaptive sampling mechanism based on feature importance. The model first performs optimal binning preprocessing on the input static features, discretizing continuous variables into high-information-gain categorical features. In each iteration, the model calculates the negative gradient of the current ensemble model: .

[0116] Subsequently, training samples are weighted based on feature importance weights, focusing on learning difficult-to-classify samples. The construction process of the base learner $h_m(x)$ adopts a feature importance-weighted splitting criterion: ; in, The sample weights are positively correlated with feature importance. This mechanism enables the model to adaptively focus on key risk features, significantly improving the accuracy and stability of identifying static risks.

[0117] Furthermore, the dynamic risk prediction sub-model employs a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN) specifically for processing time-series data such as cash flow and project progress, capturing the dynamic evolution patterns of risk. The data processing procedure for this model is as follows: The system receives normalized time-series features from multi-source data streams. First, it uses dilated causal convolution of the temporal convolutional network for hierarchical feature extraction, the operation of which can be described as follows: ; Where W is the convolution kernel weight and d is the dilation factor, thus capturing multi-scale temporal patterns. Subsequently, the output high-dimensional feature sequence is input into an LSTM unit, and its gating mechanism (input gate, forget gate, output gate) models long-term dependencies. The forget gate decision logic can be expressed as: .

[0118] This hybrid structure combines the parallel efficiency of TCN with the long-range modeling advantages of LSTM, significantly improving the accuracy and timeliness of risk trend prediction.

[0119] Furthermore, the graph-based risk transmission sub-model utilizes graph neural networks (GNNs) to learn the risk transmission effects of related entities from a dynamic knowledge graph. The data processing involves abstracting the regulated object and its related parties as nodes in a graph structure. The core of its data processing lies in hierarchical graph attention propagation. The system first calculates the adaptive attention coefficients between nodes to quantify the intensity of risk contagion. ; in, Let be the feature vector of node i, and W be the shared weight matrix. This is the attention vector. Then, by aggregating the risk information of neighboring nodes, the hidden state of the target node is updated: .

[0120] This mechanism can dynamically capture the asymmetric transmission effect of risks between nodes, thereby enabling accurate tracing of the source and prediction of the diffusion path of systemic financial risks.

[0121] Specifically, risk fusion and output: The prediction results of the above three sub-models are weighted and fused (the weights can be dynamically adjusted according to the recent performance of the models), and finally a comprehensive risk probability value between 0 and 1 is output. .

[0122] Furthermore, the core of the weighted fusion process lies in constructing a dynamic weight allocation mechanism that is negatively correlated with the real-time performance and data freshness of each sub-model. First, the system periodically evaluates the predictive performance of each sub-model within the latest time window and calculates its normalized weighted score. This score is determined by prediction accuracy. Data timeliness and result stability The calculation formula is as follows: ; in, , , Assigning fixed importance weights to each evaluation indicator. The variance of the predicted results is used to penalize instability.

[0123] Subsequently, the dynamic weights of each model in the final fusion are calculated based on this score. : ; in, This is a scaling factor used to control the degree of differentiation in weight allocation.

[0124] Ultimately, the system's comprehensive risk prediction value Calculated by the following formula: ; in, These are the prediction outputs of the dynamic, graphical, and static sub-models, respectively. This method achieves adaptive optimization of the fusion weights by dynamically evaluating model reliability and data quality, thereby ensuring that the comprehensive risk assessment is accurate, timely, and robust.

[0125] In one embodiment of the present invention, the multi-objective optimization decision engine includes: This module is responsible for finding the optimal combination of warning threshold and response strategy.

[0126] Specifically, the optimization problem modeling involves: Decision variables (X): Threshold vector T_i: Defines a set of personalized dynamic thresholds for each project i, for example: T_i = [Minimum balance threshold_θ1, Project schedule deviation warning threshold_θ2, Cash outflow speed threshold_θ3, ...]; Strategy mapping matrix S: Defines a mapping relationship from "risk level - risk type" to specific response strategies. For example, for "high risk - abnormal cash outflow", the mapped strategy combination might be [Strategy A: Upgrade the approval process for fund disbursement to the municipal regulatory agency, Strategy B: Require the submission of a funding plan for the next three months, Strategy C: Suspend non-project fund payments]; Specifically, the objective function (F) involves simultaneously optimizing three competing core objectives: F1: Minimizing the total system risk. Minimize Σ (ω_i * P_risk_i), where ω_i is the weight of the project (which can be set according to the project size and the number of homebuyers involved); F2: Maximizing the overall efficiency of escrow fund utilization. Maximize Σ (Utilization_i), where Utilization_i can be defined as (funds allocated for project construction / average escrow balance), measuring the degree of fund stagnation; F3: Maximizing the probability of on-time project delivery. Maximize Σ (P_onTime_i), this probability can be given by an independent prediction model or calculated as an intermediate result of optimization.

[0127] Specifically, the constraints (C) are as follows: Fund security bottom line constraint: The balance of the supervision account for any project shall not be lower than the legally mandated absolute minimum amount; Strategy feasibility constraint: Implementation costs of the strategy, administrative resource restrictions, etc.; Policy guidance constraint: Such as the priority adjustment of "ensuring delivery of buildings" in a specific period.

[0128] Specifically, regarding the optimization algorithm and solution: Since this optimization problem is characterized by high dimensionality, nonlinearity, and multiple objectives, we employ an improved third-generation multi-objective evolutionary algorithm—the Non-Dominated Sorting Genetic Algorithm Based on Reference Points (NSGA-III)—to solve it. NSGA-III better maintains the distribution of solutions on the Pareto front when handling optimization problems with three or more objectives.

[0129] Further, the solution process is as follows: Initialization: A population of N "individuals" is randomly generated, each representing a complete system configuration scheme (i.e., T and S for all items); Simulation and Evaluation: Based on historical data or a constructed digital twin environment, Monte Carlo simulations are performed on the scheme represented by each individual in the population to predict the operation within a future planning cycle (e.g., a quarter), thereby calculating the three objective function values ​​(F1, F2, F3) for each individual; Evolutionary Iteration: The algorithm performs selection, crossover, and mutation operations to generate a progeny population. Through the non-dominated sorting and reference point-based selection mechanism of NSGA-III, the population is guided to evolve towards the true Pareto front; Output: After multiple iterations, the algorithm outputs a "Pareto optimal solution set". Each solution in this set represents a non-dominated, excellent system configuration scheme that cannot be improved simultaneously on the three objectives.

[0130] In one embodiment of the present invention, the human-machine collaborative decision-making and strategy execution module includes: Specifically, Pareto front visualization and interaction: The system provides regulatory decision-makers with a visualization interface that displays the Pareto optimal solution set in the form of a 3D scatter plot or parallel coordinate axes. Decision-makers can clearly see the trade-offs between different options on the three objectives (e.g., how much efficiency needs to be sacrificed to improve safety by 1%).

[0131] Specifically, when using a three-dimensional scatter plot, the system optimizes three core objectives—risk control level, risk control degree, and risk control level. Liquidity of funds With regulatory costs —Directly mapped to coordinate axes in three-dimensional space. Each Pareto optimal solution It is represented as a data point in this space, and its coordinates are: .

[0132] This system innovatively introduces a solution set density coloring mechanism, where the color depth of a point is inversely proportional to the local solution set density, i.e., This allows for a more intuitive highlighting of unique solutions in sparsely distributed areas, assisting decision-makers in making diverse choices.

[0133] Furthermore, when using a parallel coordinate axis, the system maps N optimization objectives to N perpendicular and parallel axes. A solution... Represented as a broken line crossing all axes, its ordinate on the i-th axis is the normalized objective function value. The innovation of this invention lies in the introduction of interactive dimension filtering and a weight slider, allowing decision-makers to adjust the weight vector. At that time, the view highlights the solution that is closest to the ideal point under the current weight, and its utility value is given by the formula: .

[0134] Furthermore, through dynamic computation, this dual visualization mechanism transforms the complex multi-objective trade-off process into an intuitive graphical interaction, greatly improving decision-making efficiency and system transparency.

[0135] Specifically, in the final solution selection process, decision-makers can manually choose the solution that best aligns with their current management intentions from the Pareto set, based on current macroeconomic policy orientation, market environment, and regulatory priorities. For example, during periods of high risk in the real estate market, a solution more inclined towards F1 (safety) can be selected; during periods of market stability, a solution more inclined towards F2 (efficiency) can be chosen. The system also supports automatic selection based on preset rules.

[0136] Furthermore, the system will use the preference weight vector set by the decision-maker. The standardized objective vector of each solution in the Pareto solution set Matching is performed. The degree to which each solution matches the subjective intent is quantified by calculating a weighted cosine similarity. .

[0137] Furthermore, the system then automatically highlights and recommends similarity scores. The highest optimal compromise solution is found, and the corresponding threshold and strategy combination is used as the final execution plan. This process combines quantitative analysis with qualitative decision-making, realizing intelligent decision-making through human-machine collaboration.

[0138] Specifically, the automated execution of the strategy (ATO) involves the system automatically updating the new threshold vector {T_i} to the core database and deploying the strategy mapping matrix S to the workflow engine once the solution is determined. When real-time data from a project triggers a new personalized threshold, the system automatically initiates a pre-defined strategy workflow, sending instructions or tasks to relevant parties (banks, real estate companies, supervisors) via interfaces to achieve precise and rapid intervention.

[0139] In one embodiment of the present invention, the continuous learning and feedback optimization module includes: Specifically, this module enables the system to self-evolve.

[0140] Specifically, the system performs the following: Effectiveness Evaluation: It continuously tracks the actual effects of each implemented strategy, recording data throughout the entire process from early warning to risk mitigation (or deterioration); Feedback Learning: Incremental Model Updates: The risk prediction module periodically performs incremental training using new data to maintain its predictive accuracy. The process involves a preset training trigger cycle. When new regulatory data accumulates to a set time window, the incremental training process is automatically initiated. This process uses a rolling time window method, retaining key statistical features of historical data and combining them with new data to form an incremental training set. This allows for parameter fine-tuning of the existing prediction model, rather than a full reconstruction. This significantly reduces computational resource consumption and endows the model with dynamic evolution capabilities, enabling it to continuously track the time-varying nature of financial risk patterns, effectively overcoming model aging issues, and thus ensuring the timeliness and accuracy of risk warnings. Specifically, the optimizer uses reinforcement learning: the actual effect after policy execution (compared to simulated predictions) is used as a reward signal and fed back to the multi-objective optimization engine. By integrating reinforcement learning ideas (such as multi-objective multi-armed gambling machine algorithms), the system acts as an intelligent agent, treating threshold setting and policy selection as the action space. Based on historical data, it explores and utilizes trade-offs, using multi-objective reward vectors (such as risk reduction rate and capital utilization rate) to optimize the Pareto front. A scalarization function is innovatively introduced. ; Wherein, it represents the first One target reward, For adaptive weights, s is the state. This method enables threshold self-adjustment, improving the accuracy and efficiency of monitoring. The optimization engine can learn which types of (threshold, strategy) combinations are more effective in the real world, thus developing preferences in future optimized searches and gradually improving the quality of its recommendations.

[0141] This invention discloses a pre-sale fund supervision hardware based on multi-objective optimization. It achieves precise quantification of risk transmission by constructing a multi-source heterogeneous data fusion and dynamic knowledge graph, and uses ensemble learning and time-series models for dynamic risk prediction. The hardware system solves the optimal balance between fund security, liquidity efficiency, and project delivery through a multi-objective optimization decision engine, and achieves precise strategy matching and automated execution through a human-machine collaborative interface. It effectively solves the core problems of data fragmentation, delayed risk identification, and rigid decision-making in traditional supervision hardware systems, realizing closed-loop automated supervision from data perception and intelligent decision-making to strategy execution, significantly improving the overall supervision efficiency, adaptability, and risk prevention and control level of the system.

[0142] Example 3 To achieve the above invention, such as Figure 2As shown, this embodiment also provides a pre-sale fund supervision device 10 based on multi-objective optimization, the device 10 including: The Dynamic Knowledge Graph Construction Module 100 is used to construct a dynamic knowledge graph. Through the fusion and entityization of multi-source heterogeneous data, it forms an entity relationship network that includes regulated projects, real estate development companies, regulatory accounts and related parties, and quantifies the risk transmission path.

[0143] The multi-model fusion prediction module 200 is used to generate a comprehensive risk probability value for regulatory projects based on dynamic knowledge graphs and multi-source heterogeneous data, using a multi-model fusion prediction method that integrates static feature models, time series models and graph neural networks.

[0144] The multi-objective optimization decision module 300 is used to input the comprehensive risk probability value into the multi-objective optimization decision engine, solve the Pareto optimal solution set of capital security, liquidity efficiency and project delivery success rate through an improved multi-objective evolutionary algorithm, and adjust the optimization objective weights according to dynamic constraints.

[0145] The human-machine collaboration and dynamic monitoring module 400 is used to display the Pareto optimal solution set through the human-machine collaboration decision interface, match the optimal solution based on the preference weight set by the decision-maker, generate personalized early warning threshold vector and response strategy combination, and deploy it to the automated execution module to realize dynamic monitoring.

[0146] In one embodiment of the present invention, it further includes: a continuous learning and feedback optimization module, used to feed back the actual effect data after strategy execution to the risk prediction module and the multi-objective optimization decision engine, fine-tune the parameters of the prediction model based on the incremental training method, and update the optimizer through reinforcement learning using a multi-objective multi-armed gambling machine algorithm.

[0147] This invention discloses a pre-sale fund supervision device based on multi-objective optimization. It achieves precise quantification of risk transmission paths by constructing a dynamic knowledge graph and generates a comprehensive project risk probability using a multi-model fusion method. The device utilizes a multi-objective optimization decision engine to solve for the Pareto optimal solution set of fund security, liquidity efficiency, and project delivery success rate. Through a human-machine collaboration mechanism, it achieves personalized early warning and precise strategy matching. It effectively overcomes the core defects of traditional supervision systems, such as lagging risk identification, single decision-making dimensions, and lack of dynamic adjustment capabilities. It realizes fully automated supervision from risk perception and intelligent decision-making to strategy execution, significantly improving the systematicness, accuracy, and adaptability of fund supervision.

[0148] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 3As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the pre-sale fund supervision method based on multi-objective optimization described above.

[0149] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a pre-sale fund supervision method based on multi-objective optimization as described in the foregoing embodiments.

[0150] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0151] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A pre-sale fund supervision method based on multi-objective optimization, characterized in that, include: S1, construct a dynamic knowledge graph, and through the fusion and entity processing of multi-source heterogeneous data, form an entity relationship network including regulated projects, real estate development companies, regulatory accounts and related parties, and quantify the risk transmission path; S2, based on dynamic knowledge graphs and multi-source heterogeneous data, adopts a multi-model fusion prediction method that integrates static feature models, time series models and graph neural networks to generate comprehensive risk probability values ​​for regulatory projects; S3, input the comprehensive risk probability value into the multi-objective optimization decision engine, solve the Pareto optimal solution set of capital security, liquidity efficiency and project delivery success rate through the improved multi-objective evolutionary algorithm, and adjust the optimization objective weights according to dynamic constraints. S4. The Pareto optimal solution set is displayed through the human-machine collaborative decision-making interface. The optimal solution is matched based on the preference weights set by the decision-maker, and a personalized early warning threshold vector and response strategy combination are generated and deployed to the automated execution module to achieve dynamic monitoring.

2. The method as described in claim 1, characterized in that, The construction of a dynamic knowledge graph, through the fusion and entity-based processing of multi-source heterogeneous data, forms an entity relationship network including regulated projects, real estate development companies, regulatory accounts, and related parties, and quantifies risk transmission paths, including: S11, based on the preset regulatory entity definition, uses entity linking and relationship extraction technology to fuse multi-source heterogeneous data and generate a dynamic knowledge graph containing node attributes and edge relationships. S12 uses a graph computing engine to calculate node centrality, community clustering, and risk transmission paths in real time, and outputs graph feature vectors to a multi-objective optimization decision engine.

3. The method as described in claim 1, characterized in that, The method, based on dynamic knowledge graphs and multi-source heterogeneous data, employs a multi-model fusion prediction approach integrating static feature models, time-series models, and graph neural networks to generate a comprehensive risk probability value for regulatory projects, including: S21 employs an improved NSGA-III algorithm for multi-objective evolutionary optimization, using non-dominated sorting and a reference point-based selection mechanism to solve for the Pareto optimal solution set. S22, based on dynamic weight allocation formula The prediction results of the three sub-models are weighted and fused to generate a comprehensive risk probability value; among them, , The normalized weighted scores for the three sub-models are: This is a scaling factor used to control the degree of differentiation in weight allocation.

4. The method as described in claim 1, characterized in that, The process involves inputting the comprehensive risk probability value into a multi-objective optimization decision engine, using an improved multi-objective evolutionary algorithm to solve for the Pareto optimal solution set for capital security, liquidity efficiency, and project delivery success rate, and adjusting the optimization objective weights according to dynamic constraints, including: S31 uses Monte Carlo simulation to evaluate the system configuration schemes represented by each individual in the population and calculates the performance values ​​of the corresponding system configuration schemes on the three objective functions of capital security, liquidity efficiency and project delivery success rate. S32 adjusts the weights of optimization targets in real time based on preset dynamic constraints, including bottom-line constraints on capital security, constraints on strategy feasibility, and constraints on policy guidance.

5. The method as described in claim 1, characterized in that, The process involves displaying the Pareto optimal solution set through a human-machine collaborative decision-making interface, matching the optimal solution based on the decision-maker's set preference weights, generating a personalized early warning threshold vector and response strategy combination, and deploying it to the automated execution module for dynamic monitoring, including: S41 uses a three-dimensional scatter plot to display the Pareto optimal solution set, where the color depth of the points is inversely proportional to the density of the local solution set, in order to assist decision-makers in making diverse choices. S42, by calculating weighted cosine similarity Quantify the degree of alignment between each solution and the decision-maker's subjective intent, and highlight and recommend the optimal compromise solution with the highest similarity; among which, This represents the preference weight vector set by decision-makers, reflecting the degree of importance attached to different objectives. Indicates the first A vector of objective function values ​​for each Pareto solution.

6. The method as described in claim 1, characterized in that, Also includes: S5, through the continuous learning and feedback optimization module, feeds back the actual effect data after the strategy is executed to the risk prediction module and the multi-objective optimization decision engine. Based on the incremental training method, the prediction model is fine-tuned, and the optimizer is updated through reinforcement learning using the multi-objective multi-armed gambling machine algorithm.

7. A pre-sale fund monitoring device based on multi-objective optimization, characterized in that, include: The dynamic knowledge graph construction module is used to build dynamic knowledge graphs. Through the fusion and entity processing of multi-source heterogeneous data, it forms an entity relationship network that includes regulated projects, real estate development companies, regulatory accounts and related parties, and quantifies the risk transmission path. The multi-model fusion prediction module is used to generate a comprehensive risk probability value for regulatory projects based on dynamic knowledge graphs and multi-source heterogeneous data, using a multi-model fusion prediction method that integrates static feature models, time series models and graph neural networks. The multi-objective optimization decision module is used to input the comprehensive risk probability value into the multi-objective optimization decision engine, solve the Pareto optimal solution set of capital security, liquidity efficiency and project delivery success rate through an improved multi-objective evolutionary algorithm, and adjust the optimization objective weights according to dynamic constraints. The human-machine collaboration and dynamic monitoring module is used to display the Pareto optimal solution set through the human-machine collaborative decision-making interface, match the optimal solution based on the preference weight set by the decision-maker, generate personalized early warning threshold vectors and response strategy combinations, and deploy them to the automated execution module to achieve dynamic monitoring.

8. The apparatus as claimed in claim 7, characterized in that, Also includes: The continuous learning and feedback optimization module is used to feed back the actual effect data after the strategy is executed to the risk prediction module and the multi-objective optimization decision engine. It fine-tunes the parameters of the prediction model based on the incremental training method and updates the optimizer through reinforcement learning using the multi-objective multi-armed gambling machine algorithm.

9. An electronic device, comprising: processor; The memory stores executable instructions; when the processor executes the instructions, it implements the pre-sale fund supervision method based on multi-objective optimization as described in any one of claims 1-6.

10. A computer-readable storage medium storing a computer program, which, when executed by a processor, implements a pre-sale fund supervision method based on multi-objective optimization as claimed in any one of claims 1-6.

Images