A decision-making method for multi-objective optimization based on a project investment strategy model

Abstract

The application discloses a kind of based on project investment strategy model carries out multi-objective optimization decision-making method, belong to venture capital technical field, first, construct project investment field knowledge graph and as analogy complex system decision-making field knowledge graph, and carry out cross-domain atlas fusion;Second, investment strategy is converted into project portfolio decision model based on mapping rule;Then, construct time series graph neural network risk transmission quantitative evaluation model, dynamically calculate the risk transmission coefficient between projects, to identify previously unseen "emergent risk";Next, using multi-objective optimization algorithm, generate multiple candidate project investment portfolio scheme;Again, the scheme is checked and the market adaptability is verified between projects;Finally, output recommended scheme, and continuously optimize the model according to the post-investment performance of the project through reinforcement learning mechanism.The application can obtain the investment decision-making effect that the prior art cannot reach, and the yield and system robustness are considered.

Description

Technical Field

[0001] This invention belongs to the field of venture capital technology, and more specifically, relates to a decision-making method based on a project investment strategy model for multi-objective optimization. Background Technology

[0002] Private equity funds typically follow a rigorous process when investing in projects, including project sourcing, initial screening, due diligence, investment decision-making, post-investment management, and exit. This process generates a massive amount of structured data (such as financial statements) and unstructured data (such as business plans, due diligence reports, industry research reports, and expert interview transcripts).

[0003] Existing technologies for processing this data to aid investment decisions suffer from the following limitations: whether employing traditional financial models (such as DCF valuation) or introducing artificial intelligence for single-point analysis (such as using NLP techniques to analyze contract terms), they essentially involve isolated and atomized assessments of projects. These methods can effectively evaluate the financial health and market prospects of individual projects, but they completely fail to identify and quantify the "structural risks" that emerge when a portfolio of multiple projects is considered as a whole. For example, systemic risks such as implicit technological dependencies, supply chain overlaps, competition for key talent, or potential market competition between projects are invisible under current technological frameworks because they are not documented in individual project documents but exist within a complex network of relationships between projects. This means that investment decisions may seem reasonable at the micro level (individual projects) but harbor significant hidden dangers at the macro level (the entire portfolio). How to effectively identify, quantify, and mitigate these structural risks remains a long-standing and unresolved technical challenge. Summary of the Invention

[0004] To address the above shortcomings, this invention provides a decision-making method for multi-objective optimization based on a project investment strategy model, comprising the following steps:

[0005] S1. Extract entities and relationships from structured and unstructured data throughout the entire project investment process and store them as a time-stamped knowledge sub-graph of the project investment domain;

[0006] S2. Construct an analogous domain knowledge subgraph for decision-making in complex systems;

[0007] S3. Perform cross-domain knowledge graph fusion and semantic association on the knowledge subgraph of the project investment domain and the knowledge subgraph of the analogy domain, and establish a mapping relationship rule set between the two graph nodes, including semantic layer alignment and graph structure fusion. Based on the rule set, generate a unified fusion graph view containing cross-domain ontology concepts, wherein the entities in the investment domain inherit the computational attributes and constraint logic of the corresponding entities in the complex system domain.

[0008] S4. Based on the mapping relationship rule set, the project investment strategy is structured into a multi-objective project portfolio decision model that includes optimization objectives, decision variables and constraints, wherein the business constraints in the strategy are automatically translated into the systems engineering constraints in the fusion graph view;

[0009] S5. Based on the knowledge subgraph of the project investment field, construct a risk transmission quantitative assessment model and dynamically calculate the risk transmission coefficient between projects;

[0010] S6. Based on the multi-objective project portfolio decision model generated in step S4, the risk transmission coefficient in step S5, and the project investment domain knowledge sub-graph, and using the project feature data to calculate the objective function of the multi-objective optimization algorithm, the multi-objective optimization algorithm is used to search for project portfolios and generate multiple project portfolio schemes.

[0011] S7. Based on the constraints in the multi-objective project portfolio decision-making model generated in S4, and combined with the market conditions, verify the rationality of the decision-making of the multiple project portfolio schemes generated in S6.

[0012] S8. Output the validated project portfolio plan, and dynamically optimize the risk transmission quantitative assessment model of S5 and the target weights of the multi-objective optimization algorithm of S6 based on the actual performance of the projects after investment through reinforcement learning mechanism.

[0013] Furthermore, the types of entities in step S1 include one or more of the following: project or target company, founding team or key personnel, core technology, product or service, market segmentation, upstream and downstream of the industrial chain, competitors, risk events, and financial indicators.

[0014] Further, step S3 includes:

[0015] S301. Semantic layer alignment: Use a pre-trained language model to calculate the cosine similarity of word vectors of core words across domains, and perform syntactic dependency analysis and semantic role labeling.

[0016] S302. Graph structure fusion: A graph matching network is used to learn the joint features of two knowledge sub-graphs, and a set of mapping relationship rules between the nodes of the two graphs is established based on the structural similarity score.

[0017] Further, step S4 includes:

[0018] S401. Based on the mapping relationship rule set, the input text-based project investment strategy is parsed into a structured demand vector;

[0019] S402. Based on the structured demand vector and the mapping relationship rule set, automatically generate the optimization objective, decision variables and constraints of the multi-objective project combination decision model.

[0020] Furthermore, in step S5, the risk transmission quantification assessment model is a time-series graph neural network model, and its risk transmission coefficient calculation formula is as follows:

[0021] ;

[0022] In the above formula, This indicates the intensity of risk transmission from project i to project j at time t;

[0023] This indicates the degree of relevance of project i and project j in the industrial chain and technology ecosystem, taking into account whether there are upstream and downstream supply relationships, technological dependence relationships or direct competitive relationships between the projects.

[0024] Indicates the time decay factor;

[0025] This represents the market volatility coefficient, which has a non-linear interaction with the inherent risks between projects;

[0026] Indicates model parameters.

[0027] Further, in step S6, the multi-objective optimization algorithm is a multi-objective particle swarm optimization algorithm, and its objective function includes:

[0028] Maximize returns: ;

[0029] Minimize risk: ;

[0030] Maximize strategic synergy: ;

[0031] In the above formula, These are the investment weights for projects i and j, respectively. Let be the expected internal rate of return for project i. Let the revenue covariance of projects i and j be... The comprehensive risk transmission coefficient of project i is calculated from the risk transmission quantitative assessment model; The score is given for the synergistic effect between items i and j, calculated based on the knowledge graph.

[0032] Further, step S7 includes:

[0033] S701. Constraint check: Based on the constraints generated in step S4, check whether each candidate combination meets all hard requirements.

[0034] S702. Project portfolio synergy or conflict check: Based on knowledge graph query, check whether there are conflicts of interest, overlapping technology paths or industrial chain synergy effects in the project portfolio.

[0035] S703. Market adaptability assessment, including using support vector machines to classify market conditions, constructing a vector autoregression model to analyze the impact of macroeconomic indicators, and conducting Monte Carlo simulations to assess the returns and risks of the portfolio under different scenarios.

[0036] Furthermore, the reinforcement learning mechanism specifically employs the Q-learning algorithm;

[0037] The state represents the current market classification and combination characteristics, while the action represents adjusting the risk transmission parameters of the risk transmission model or the target weights of the multi-objective optimization algorithm.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] By analogying the project portfolio ecosystem to a complex system and introducing the knowledge system of complex systems decision theory, a brand-new investment decision analysis paradigm is constructed through cross-domain knowledge graph fusion and semantic association. Under this paradigm, fuzzy business risks are transformed into computable and inferable system structural features, enabling us to use tools such as time series graph neural networks to quantify and predict previously invisible "emergent risks".

[0040] This management approach fundamentally changes the way problems are defined and analyzed, enabling subsequent multi-objective optimization to be carried out in a new and more comprehensive dimension, thereby achieving investment decision-making results that are unattainable with existing technologies, balancing profitability and system robustness. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0042] Example

[0043] This embodiment provides a decision-making method for multi-objective optimization based on a project investment strategy model, including:

[0044] S1. Construction of a knowledge subgraph in the project investment domain (This step aims to integrate the fragmented, massive data surrounding project investment into a structured, computable knowledge network)

[0045] In detail, S101, data acquisition and preprocessing:

[0046] We collect structured and unstructured text data from the entire private equity investment process. The structured and unstructured text data are acquired from an integrated investment management system via API interfaces and database ETL tools. The investment management system includes modules for fund financial and term data, structured data for fund and project processes, investment performance data, cash flow data, and unstructured text data acquisition. The specific data acquisition process for each module is as follows:

[0047] Fund Financial and Terms Data Module: Calls the "Fund-Related Information Entry" function of the investment management system, and obtains the total amount of subscribed capital, the progress of paid-in capital, the structure of investors, the profit distribution records and cash flow data of the fund by parsing the forms of "Investor Capital Contribution", "Fund Allocation", "Fund Subscribed Capital" and "Fund Paid-in Capital". It also extracts the core agreements of the fund such as the lock-up period, performance fee terms and management fee rate from the fund establishment information and the uploaded "Limited Partnership Agreement (LPA)" document.

[0048] The Fund and Project Process Structured Data Module synchronizes the "Enter Fund Process Data" and "Enter Project Process Data" functions in the investment management system with the process status, date, decision results, and participant information for each stage (such as research and negotiation, project initiation, due diligence, investment decision, execution, and post-investment) recorded at each stage. It also obtains equity structure data such as the fund's investment amount and shareholding ratio for specific projects through the "Investment Project Data Entry" task.

[0049] Investment performance data module: Regularly captures standardized form data submitted by the "Fund Post-Investment Quarterly Report" and "Project Post-Investment Quarterly Report" functions in the investment management system, and extracts key performance and operating indicators including fund net asset value (NAV), internal rate of return (IRR), distributed income to paid-in capital multiple (DPI), residual value to paid-in capital multiple (RVPI), project operating income, net profit, etc.

[0050] Cash flow data module: It connects to the "Funds and Equity and Changes" and "Exit Execution_Cash Flow" functions under the "Post-Investment Management" stage of the investment management system, and collects the amount, date and counterparty information of all cash flow events that occur throughout the entire life cycle of the project, such as investment payments, operating income, expenses, and exit distribution.

[0051] Unstructured text data acquisition module: Scans and downloads document attachments uploaded by the investment management system at various business stages. This mainly includes:

[0052] Business Plan (BP) obtained through the "Project Reserve" module;

[0053] The "Enter Fund Process Data" and "Enter Project Process Data" functions are used to obtain the process documents uploaded at each stage. The process documents uploaded at each stage include: "Private Equity Fund Research Report" / "Negotiation Minutes" (research and negotiation stage), "Project Preliminary Review Report" / "Project Establishment Report" (project establishment stage), "Due Diligence Report" (legal, financial, and business due diligence, due diligence stage), "Investment Decision Committee Meeting Minutes" and "Review Opinions" (each investment decision stage), "Limited Partnership Agreement (LPA)", and "Fund Custody Agreement" (execution stage).

[0054] Post-investment operation reports (quarterly / annual reports) and interim tracking minutes submitted through the "Fund Post-Investment Quarterly Report", "Project Post-Investment Quarterly Report" and "Latest Project Progress" functions;

[0055] The GP / LP background information obtained through the "Partner Management" module;

[0056] The Project Exit Summary Report obtained through the "Exit Execution" phase of the project.

[0057] S102, Entity and Relation Extraction:

[0058] Information extraction is performed by using a pre-trained language model (such as FinBERT) fine-tuned with corpus data from the private equity investment field, combined with a specific model (BiLSTM-CRF model in this embodiment).

[0059] Entity Extraction: A BiLSTM-CRF model is used to extract entities from documents such as the "Business Plan" and "Due Diligence Report." Entity types primarily include: project or target company (e.g., "XX Semiconductor"), founding team or key personnel (e.g., "Zhang San (CEO)"), core technologies (e.g., "5nm chip etching technology"), products or services (e.g., "XinChi Series Automotive Chips"), market segmentation (e.g., "New Energy Vehicle Intelligent Cockpit Market"), upstream and downstream of the industry chain (e.g., "Supplier: XX Photoresist Company"), competitors (e.g., "YY Technology"), risk events (e.g., "Founder's Departure Risk"), and financial indicators (e.g., "Operating Revenue").

[0060] Relationship Extraction: A graph convolutional network (GAT) model based on attention mechanism is used to extract relationships between entities. For example: "Zhang San" - [works for] - "XX Semiconductor"; "XX Semiconductor" - [possesses] - "5nm chip etching technology"; "XX Semiconductor" - [competes with] - "YY Technology".

[0061] S103. Knowledge Graph Storage and Temporal Sequencing:

[0062] The extracted entities and relations (triples) are stored in the Neo4j graph database, with each entity as a node and each relation as an edge. A timestamp attribute is attached to each node and edge to record the effective time of the knowledge fact, forming a knowledge sub-graph of the project investment domain.

[0063] S2. Construction of an analogy domain knowledge subgraph (In order to understand the systematic risk of the portfolio from a new perspective, this step introduces the "complex systems decision-making domain" as an analogy)

[0064] In detail, S201, Knowledge Acquisition: Extract core concepts from publicly available Project Management Body of Knowledge (PMBOK), game theory, and systems engineering literature, such as decision-making entities, resource pools, resource constraints, task dependencies, and conflict resolution mechanisms.

[0065] S202. Graph Construction: Use the Web Ontology Language (OWL) to construct a subgraph of knowledge analogous to the domain, where nodes represent decision-making entities, resources, and tasks, and edges represent relationships such as decision dependencies, resource consumption, and temporal constraints.

[0066] S3. Cross-domain knowledge graph fusion and semantic association (This step is the core technology of this invention, aiming to establish a connection between the investment field and the complex systems field. It includes semantic layer alignment and graph structure fusion)

[0067] In detail, S301, semantic layer alignment: The FinBERT model is used to calculate the cosine similarity of word vectors of core words in two domains (such as "investment portfolio" in the investment field and "multi-task scheduling" in the complex systems field) to establish semantic mapping.

[0068] S302, Graph Structure Fusion: A Graph Matching Network (GMN) is used to jointly learn features from two knowledge sub-graphs, and a set of mapping rules between nodes is established based on structural similarity scores. For example, the rules are: mapping "projects" in the portfolio to "tasks" in a complex system; mapping "investment capital" to "resource pools"; and mapping "competitive relationships between projects" to "resource competition between tasks".

[0069] S4. Decision Model Structuring (Based on the mapping relationship rule set established in S3, unstructured investment strategies are automatically transformed into computable decision models).

[0070] In detail, S401, Demand Analysis: Analyze textual investment strategies (such as "finding a portfolio of stable growth projects in the TMT industry with moderate risk and an expected exit within 5 years") into structured demand vectors.

[0071] S402. Model Generation: Based on demand vectors and mapping rules, automatically generate the optimization objectives, decision variables, and constraints of a multi-objective project portfolio decision-making model. The constraints are the portfolio construction rules generated according to the investment strategy, such as the upper limit of total investment, the investment ratio of a single project, or the industry allocation ratio.

[0072] S5. Construction of a Quantitative Assessment Model for Risk Transmission (This step aims to construct a quantitative assessment model for risk transmission to quantify the "emergent risks" between projects. In this embodiment, the model is specifically implemented by constructing a Temporal Graphical Neural Network (T-GNN) model.)

[0073] Detailed dynamic risk transmission modeling: Construct a time-series graphical neural network (T-GNN) model to dynamically calculate the risk transmission strength between projects. The risk transmission coefficient formula is as follows:

[0074] ;

[0075] In the above formula, This indicates the intensity of risk transmission from project i to project j at time t;

[0076] This indicates the degree of relevance between project i and project j in the industrial chain and technology ecosystem. It considers whether there are upstream and downstream supply relationships, technological dependence relationships or direct competitive relationships between projects. That is, it is calculated in the knowledge graph by a meta-path-based algorithm. If project i and project j share upstream suppliers, rely on the same technology, or target the same market segment, the value is higher.

[0077] Indicates the time decay factor;

[0078] It represents the market volatility coefficient, which has a non-linear interaction with the inherent risks between projects. That is, when the market fluctuates sharply, the risk transmission effect will be significantly amplified.

[0079] Indicates model parameters.

[0080] S6. Project portfolio search based on multi-objective optimization: Use the multi-objective particle swarm optimization algorithm (MOPSO) to search for the Pareto optimal portfolio among all candidate projects.

[0081] In detail, the objective function includes:

[0082] Maximize returns: ;

[0083] Minimize risk: ;

[0084] Maximize strategic synergy: ;

[0085] In the above formula, These are the investment weights for projects i and j, respectively. Let be the expected internal rate of return for project i. Let the revenue covariance of projects i and j be... The comprehensive risk transmission coefficient of project i is calculated from the risk transmission quantitative assessment model; The score is given for the synergistic effect between items i and j, calculated based on the knowledge graph.

[0086] Using the specific values ​​extracted and quantified from the knowledge graph in the previous steps (such as the expected internal rate of return of each item extracted and quantified from the knowledge graph) , revenue covariance Synergistic effect score (etc.), as input to the multi-objective particle swarm optimization algorithm (MOPSO), the algorithm uses the investment weights of each project. Using the above input data as decision variables, the scores of three objective functions (return, risk, and synergy) are calculated, and through iterative search, a series of non-dominated solutions that achieve a balance on different objectives are finally generated, namely multiple project portfolio schemes with different focuses.

[0087] S7. Verification of decision rationality (conducting feasibility and robustness tests on the candidate combination schemes generated in S6).

[0088] In detail, S701, constraint check: Based on the constraints generated in step S4 (such as the upper limit of total investment and industry allocation restrictions), check whether each candidate combination meets all the hard requirements;

[0089] S702. Project Portfolio Collaboration or Conflict Check: The Cypher Graph Query Language is used to automatically check for conflicts within the merged knowledge graph. For example, executing the query: "MATCH (p1:Project)-[:COMPETES_WITH]->(p2:Project)WHERE p1.name IN [Combination List] AND p2.name IN [Combination List] RETURN p1, p2", if results are returned, it indicates a direct competition conflict within the portfolio.

[0090] S703. Market Adaptability Assessment: Support Vector Machine (SVM) is used to classify the current macro market state (bull market / bear market / sideways market), and combined with Vector Autoregression (VAR) model and Monte Carlo simulation, the performance of the portfolio under different market scenarios is evaluated.

[0091] S8: Decision Output and Feedback Optimization (The optimal project portfolio solution, after verification and ranking, will be recommended to the investment decision committee through a visual interface)

[0092] Feedback and optimization mechanism: Collect data on the final decision of the investment committee and the actual performance of the project after investment.

[0093] Using the Q-learning reinforcement learning algorithm, (market state, portfolio features) is taken as the state, and (adjusting risk model parameters or optimizing target weights) is taken as the action. The reward function R is defined as the risk-adjusted return deviation, that is, when the actual Sharpe ratio of the portfolio is higher than the model prediction value, a positive reward is obtained, and vice versa. Through this mechanism, the system can continuously iterate and optimize the model parameters in S5 and S6, thereby continuously improving the decision-making ability.

[0094] It should be noted that the structure of the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Any equivalent modifications made by those skilled in the art based on the content of this specification, or direct or indirect applications in other related technical fields, such as the loading and unloading of other items, are included within the protection scope of this invention.

Claims

1. A decision-making method based on a project investment strategy model for multi-objective optimization, characterized in that, Includes the following steps: S1. Extract entities and relationships from structured and unstructured data throughout the entire project investment process and store them as a time-stamped knowledge sub-graph of the project investment domain; S2. Construct an analogous domain knowledge subgraph; S3. Perform cross-domain knowledge graph fusion and semantic association on the knowledge subgraph of the project investment domain and the knowledge subgraph of the analog domain, and establish a set of mapping relationship rules between the nodes of the two graphs, including semantic layer alignment and graph structure fusion; S4. Based on the mapping relationship rule set, the project investment strategy is structured into a multi-objective project portfolio decision model that includes optimization objectives, decision variables and constraints; S5. Based on the knowledge subgraph of the project investment field, construct a risk transmission quantitative assessment model and dynamically calculate the risk transmission coefficient between projects; S6. Based on the multi-objective project portfolio decision model generated in step S4, the risk transmission coefficient in step S5, and the project investment domain knowledge sub-graph, and using the project feature data to calculate the objective function of the multi-objective optimization algorithm, the multi-objective optimization algorithm is used to search for project portfolios and generate multiple project portfolio schemes. S7. Based on the constraints in the multi-objective project portfolio decision-making model generated in S4 (such as the upper limit of total investment and industry allocation restrictions), and combined with the market conditions, verify the rationality of the decision-making of multiple project portfolio schemes generated in S6. S8. Output the validated project portfolio plan, and dynamically optimize the risk transmission quantitative assessment model of S5 and the target weights of the multi-objective optimization algorithm of S6 based on the actual performance of the projects after investment through reinforcement learning mechanism.

2. The decision-making method for multi-objective optimization based on a project investment strategy model as described in claim 1, characterized in that: The types of entities in step S1 include one or more of the following: project or target company, founding team or key personnel, core technology, product or service, market segmentation, upstream and downstream of the industrial chain, competitors, risk events, and financial indicators.

3. The decision-making method for multi-objective optimization based on a project investment strategy model as described in claim 1, characterized in that: Step S3 includes: S301. Semantic layer alignment: Use a pre-trained language model to calculate the cosine similarity of word vectors of core words across domains, and perform syntactic dependency analysis and semantic role labeling. S302. Graph structure fusion: A graph matching network is used to learn the joint features of two knowledge sub-graphs, and a set of mapping relationship rules between the nodes of the two graphs is established based on the structural similarity score.

4. The decision-making method for multi-objective optimization based on a project investment strategy model as described in claim 1, characterized in that: Step S4 includes: S401. Based on the mapping relationship rule set, the input text-based project investment strategy is parsed into a structured demand vector; S402. Based on the structured demand vector and the mapping relationship rule set, automatically generate the optimization objective, decision variables and constraints of the multi-objective project combination decision model.

5. The decision-making method for multi-objective optimization based on a project investment strategy model as described in claim 1, characterized in that: In step S5, the risk transmission quantitative assessment model is a time-series graph neural network model, and its risk transmission coefficient calculation formula is as follows: ； In the above formula, This indicates the intensity of risk transmission from project i to project j at time t; This indicates the degree of relevance of project i and project j within the industry chain and technology ecosystem; Indicates the time decay factor; Indicates the market volatility coefficient; Indicates model parameters.

6. The decision-making method for multi-objective optimization based on a project investment strategy model as described in claim 5, characterized in that: In step S6, the multi-objective optimization algorithm is a multi-objective particle swarm optimization algorithm, and its objective function includes: Maximize returns: ; Minimize risk: ; Maximize strategic synergy: ; In the above formula, These are the investment weights for projects i and j, respectively. Let be the expected internal rate of return for project i. Let the revenue covariance of projects i and j be... This represents the comprehensive risk transmission coefficient of project i, calculated from the risk transmission model. The score is given for the synergistic effect between items i and j, calculated based on the knowledge graph.

7. The decision-making method for multi-objective optimization based on a project investment strategy model as described in claim 1, characterized in that: Step S7 includes: S701. Constraint check: Based on the constraints generated in step S4, check whether each candidate combination meets all hard requirements. S702. Project portfolio synergy or conflict check: Based on knowledge graph query, check whether there are conflicts of interest, overlapping technology paths or industrial chain synergy effects in the project portfolio. S703. Market adaptability assessment, including using support vector machines to classify market conditions, constructing a vector autoregression model to analyze the impact of macroeconomic indicators, and conducting Monte Carlo simulations to assess the returns and risks of the portfolio under different scenarios.

8. The decision-making method for multi-objective optimization based on a project investment strategy model as described in claim 1, characterized in that: The reinforcement learning mechanism specifically employs the Q-learning algorithm; The state represents the current market classification and combination characteristics, while the action represents adjusting the risk transmission parameters of the risk transmission model or the target weights of the multi-objective optimization algorithm.

Images