Investment and financing project recommendation method based on dynamic knowledge graph and large language model

Abstract

The invention relates to an investment and financing project recommendation method based on a dynamic knowledge graph and a large language model. The method comprises the following steps: constructing a micro knowledge graph in the field of investment and financing; performing semantic alignment on the portrait information of the target investment institution and the micrograph of knowledge, determining a corresponding node of the portrait information in the graph, and taking the node as an initial node of graph reasoning; taking a starting node as an original point, performing multiple rounds of message passing and neighborhood information aggregation on the investment and financing domain knowledge graph by adopting a graph neural network model, and calculating the importance weight of neighborhood node information by utilizing an embedded vector of a relation type in a message passing process; generating a semantic representation vector of each candidate project node after multiple rounds of iteration; and calculating correlation scores of the candidate items and the starting node based on the semantic representation vector to generate a sorting list, and decoding a key path activated in the process of the graph neural network model as an interpretable context. By adopting the technical scheme, the accuracy of project recommendation can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of natural language processing technology, and in particular to a method for recommending investment and financing projects based on dynamic knowledge graphs and large language models. Background Technology

[0002] With the continued surge in global innovation and entrepreneurship, the information in the investment and financing market is experiencing explosive growth. Investment institutions face the enormous challenge of efficiently identifying and evaluating early-stage or growth-stage projects with high growth potential from massive amounts of multi-source, heterogeneous, and unstructured business information.

[0003] Traditional methods for recommending and discovering investment and financing projects mainly rely on manual research, industry seminars, and intermediary recommendations, which suffer from problems such as low efficiency, strong subjectivity, and narrow coverage.

[0004] In recent years, with the development of information technology, some technology-enhanced solutions have emerged, but they still have significant limitations.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this application, and therefore may contain information that is not part of the prior art known to those skilled in the art. Summary of the Invention

[0006] In view of this, the present disclosure provides a method for recommending investment and financing projects based on dynamic knowledge graphs and large language models, which can improve the accuracy, interpretability, and effectiveness of decision support for project recommendations.

[0007] In a first aspect, embodiments of this disclosure provide a method for recommending investment and financing projects based on dynamic knowledge graphs and large language models, including: Construct a differentiable knowledge graph in the field of investment and financing, wherein each entity node and relation type in the differentiable knowledge graph is parameterized as an embedding vector; Semantically align the profile information of the target investment institution with the differentiable knowledge graph to determine the corresponding node of the profile information in the graph, and use it as the starting node for graph reasoning; Using the starting node as the origin, a graph neural network model is used to perform multiple rounds of message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field. The message passing process uses the embedding vector of the relation type to calculate the importance weight of the neighbor node information, and generates the semantic representation vector of each candidate project node after multiple rounds of iteration. Based on the semantic representation vector, the relevance score between the candidate items and the starting node is calculated to generate a ranking list. The key path activated during the graph neural network model is decoded as interpretability context and input into the domain expert model to generate corresponding recommendation reasons and scores.

[0008] Optionally, the construction of a differentiable knowledge graph in the investment and financing field includes: Define the ontology structure of the investment and financing field, and determine the set of entity types and relation types in the differentiable knowledge graph; Investment and financing facts are extracted from multi-source data related to the investment and financing field, forming a set of triplets with (head entity, relationship, tail entity) as the unit, and an initial symbolic knowledge graph is constructed. For each entity and each relation type in the symbolic knowledge graph, a high-dimensional distributed vector representation is assigned as the embedding vector; The embedding vectors of all assigned entities and relations are organized together into a trainable parameter matrix, and this parameter matrix is ​​registered as trainable parameters in the embedded training model. Through the vector query interface, the embedding training model can retrieve the corresponding trainable embedding vectors from the parameter matrix by accessing the unique identifiers of entities and relations, and use these vectors as model inputs to participate in forward computation and backpropagation gradient updates, thereby obtaining the differentiable knowledge graph of the investment and financing field.

[0009] Optionally, the method for recommending investment and financing projects meets one or more of the following criteria: Knowledge graph embedding technology is used to pre-train on the set of triples of the symbolic knowledge graph to obtain a semantically discriminative vector as the initial value; wherein, the knowledge graph embedding technology is a translation distance-based model, and its pre-training objective function is to minimize the distance between the operation result of the head entity vector, relation vector and tail entity vector in an effective triple and the tail entity vector. When new triples are extracted from multi-source data, embedding vectors are dynamically assigned to the new entities and relations and added to the trainable parameter matrix. At the same time, the gradient calculation and optimization range of the parameter matrix is ​​extended to these new parameters.

[0010] Optionally, the step of semantically aligning the profile information of the target investment institution with the differentiable knowledge graph to determine the corresponding node of the profile information in the graph and using it as the starting node for graph reasoning includes: Receive profile information of the target investment institution, the profile information including the institution's name text and / or preference description text; The profile information is semantically aligned with the differentiable knowledge graph. Based on the completeness of the profile information, when the institution name in the profile information can uniquely correspond to an existing entity in the differentiable knowledge graph, entity linking technology is used to map the name of the target investment institution to the corresponding specific entity node in the differentiable knowledge graph. When the profile information cannot explicitly correspond to an existing entity in the differentiable knowledge graph, dynamic node generation is performed. A virtual entity node is synthesized in the embedding space of the differentiable knowledge graph using semantic encoding and vector fusion technology. The specific entity node or the virtual entity node is determined as the starting node of the graph reasoning.

[0011] Optionally, the step of using the starting node as the origin and employing a graph neural network model to perform multi-round message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field includes: Obtain the initial feature vectors, relation type embedding vectors, and the starting node for all nodes in the differentiable knowledge graph; An iterative process of multi-round message passing and information aggregation is executed. In each iteration, for each current node in the differentiable knowledge graph, the following steps are performed: For each neighbor node of the current node, based on the relationship type connecting the neighbor node and the current node, a corresponding relationship embedding vector is selected, and the feature vector of the neighbor node is linearly transformed or nonlinearly mapped using the relationship embedding vector to generate a directed message from the neighbor node to the current node; Based on the feature vector of the current node, the feature vectors of the neighbor nodes, and the relationship embedding vector connecting the two, the attention weight of the directed message is calculated through an attention mechanism; The directed messages from all neighbor nodes are weighted and summed according to their corresponding attention weights to obtain the aggregated neighbor message of the current node; The feature vector of the current node itself is combined with the aggregated neighbor message of the current node, and a new feature vector of the current node in the next iteration is generated through an update function; After multiple rounds of preset iterations, the final updated feature vector of each node is output as its semantic representation vector.

[0012] Optionally, the step of calculating the relevance score between candidate items and the starting node based on the semantic representation vector to generate a ranked list, and decoding the key path activated during the inference process of the graph neural network model as interpretability context, inputting it into the domain expert model, and generating corresponding recommendation reasons and scores includes: Calculate the similarity between the semantic representation vector of the starting node and the semantic representation vector of each candidate item node to obtain the relevance score of each candidate item, and sort all candidate items in descending order based on the score to generate a sorted list of items. For candidate projects ranked above a predetermined threshold in the project ranking list, the information propagation path with the highest attention weight in the multi-round message passing process of the graph neural network model is traced back, and the path sequence composed of nodes and relationships is extracted. The node and relation names in the extracted path sequence are linearized into natural language descriptions according to their logical order in the path, forming a structured interpretable context. The interpretability context, basic information of candidate projects, and task instructions are combined to construct prompt words, which are then input into the domain expert model to generate recommendation reasons and scores for each candidate project.

[0013] Alternatively, the methods for recommending investment and financing projects may also include: Obtain user feedback on the project ranking list and recommendation reasons, and convert it into a monitoring signal; The gradient backpropagation algorithm is used to optimize the embedding vector of the differentiable knowledge graph, the model parameters of the graph neural network, and the generation strategy of the domain expert model using the supervision signal, and outputs the optimized final project ranking list and recommendation reasons.

[0014] Optionally, the domain expert model can be obtained in the following manner: Collect textual data in the investment and financing field and construct a fine-tuned sample set of instruction-answer pairs; Using low-rank adaptive technology, and based on the fine-tuning sample set, with the training objective of maximizing accuracy and logical consistency in investment and financing-related tasks, the natural language model is fine-tuned to obtain a domain expert model for investment and financing.

[0015] Secondly, embodiments of this disclosure provide an investment and financing project recommendation system based on dynamic knowledge graphs and large language models, comprising: A construction unit is used to construct a differentiable knowledge graph in the field of investment and financing, wherein each entity node and relation type in the differentiable knowledge graph is parameterized as an embedding vector. The alignment unit is used to semantically align the profile information of the target investment institution with the differentiable knowledge graph, determine the corresponding node of the profile information in the graph, and use it as the starting node for graph reasoning. The iterative unit is used to perform multiple rounds of message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field with the starting node as the origin and a graph neural network model. The message passing process uses the embedding vector of the relation type to calculate the importance weight of the neighbor node information, and generates the semantic representation vector of each candidate project node after multiple rounds of iteration. The recommendation unit is used to calculate the relevance score between the candidate items and the starting node based on the semantic representation vector to generate a ranked list, and decode the key path activated in the process of the graph neural network model as interpretability context, input it into the domain expert model, and generate corresponding recommendation reasons and scores.

[0016] Thirdly, embodiments of this disclosure provide an electronic device, including a memory and a processor. The memory stores a computer program executable on the processor. When the processor runs the computer program, it performs the steps of the investment and financing project recommendation method based on dynamic knowledge graphs and large language models described in any of the foregoing embodiments. Compared with the prior art, the technical solution disclosed herein has the following advantages: The investment and financing project recommendation method based on dynamic knowledge graphs and large language models provided in this disclosure parameterizes the knowledge graph into trainable embedding vectors, enabling the originally symbolic static knowledge to participate in gradient descent optimization, thereby deeply integrating domain knowledge into the model's learning process. Furthermore, by utilizing graph neural networks to perform relationship-aware message passing and reasoning on differentiable graphs, it is possible to capture the complex and deep semantic relationships between investment institutions and projects, rather than simple keyword matching. Finally, by decoding the reasoning path and integrating domain expert models to generate recommendation reasons and scores, the logic of investment and financing decisions can be understood in essence, thereby significantly improving the accuracy, interpretability, and effectiveness of project recommendations and decision support. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an investment and financing project recommendation method based on dynamic knowledge graph and large language model in one embodiment; Figure 2 This is a flowchart of a method for constructing a differentiable knowledge graph in the field of investment and financing, as shown in one embodiment. Figure 3 This is a flowchart of a semantic alignment method in one embodiment; Figure 4 This is a schematic diagram of the structure of a natural language-based intelligent item matching and recommendation system in one embodiment; Figure 5 This is a schematic diagram of the hardware structure of an electronic device in one embodiment. Detailed Implementation

[0019] As can be seen from the background technology, with the development of information technology, some technology-enhanced solutions have emerged, but they still have significant limitations, specifically: Knowledge graphs are mostly static or updated in batches on a regular basis, meaning they are static and their construction is lagging. This makes it impossible to capture market dynamics, public opinion changes, technological breakthroughs, and sudden events (such as policy releases and major collaborations) in real time, which can easily lead to outdated recommended project information or insensitivity to risks.

[0020] Traditional solutions often use rules or traditional NLP models (such as BERT+CRF) to extract entities and relations, which have limited understanding of complex semantics and unstructured text.

[0021] Specifically, traditional solutions typically rely on large amounts of labeled data, resulting in high construction and maintenance costs. Furthermore, the extracted knowledge is often static and isolated fragments, making it difficult to form a dynamic and comprehensive picture of the industry's upstream and downstream sectors, competitive landscape, and technological evolution path. When processing highly specialized, diverse, and complex texts with implicit relationships in the investment and financing field, both accuracy and recall face bottlenecks. Consequently, they fail to understand deep semantic information such as the project's business model, technological barriers, and market potential, leading to crude and poorly relevant recommendation results.

[0022] Traditional solutions are mostly based on collaborative filtering or content matching, lacking structured reasoning capabilities. Furthermore, newly registered investment institutions or newly released projects lack historical data, making it difficult to achieve effective recommendations.

[0023] To address at least one of the aforementioned technical problems, embodiments of this disclosure construct a differentiable knowledge graph and perform end-to-end joint optimization of the differentiable knowledge graph with graph neural networks and domain expert models, thereby realizing a closed-loop intelligent recommendation scheme.

[0024] Specifically, by parameterizing the knowledge graph into trainable embedding vectors, the originally symbolic static knowledge can participate in gradient descent optimization, thereby deeply integrating domain knowledge into the model's learning process. Furthermore, by using graph neural networks to perform relation-aware message passing and reasoning on differentiable graphs, the complex and deep semantic relationships between investment institutions and projects can be captured, rather than simple keyword matching. Finally, by decoding the reasoning path and integrating domain expert models to generate recommendation reasons and scores, the logic of investment and financing decisions can be understood from the essence, thereby significantly improving the accuracy, interpretability, and effectiveness of project recommendations and decision support.

[0025] To enable those skilled in the art to better understand and implement this disclosure, the following detailed description of the specific solutions, principles, advantages, and effects of this disclosure is provided with reference to the accompanying drawings and specific embodiments.

[0026] See Figure 1 , Figure 1 This is a flowchart of an investment and financing project recommendation method based on dynamic knowledge graphs and large language models, as described in this disclosure. It may include the following steps: S101, construct a differentiable knowledge graph in the field of investment and financing, wherein each entity node and relation type in the differentiable knowledge graph is parameterized as an embedding vector.

[0027] In some embodiments, a differentiable knowledge graph has been established for the investment and financing field. A differentiable knowledge graph refers to a knowledge graph that is no longer a static database, but an intelligent knowledge base that can adaptively evolve according to specific task objectives (such as precise recommendations).

[0028] In other words, the differentiable knowledge graph in this embodiment adopts an event-triggered and real-time update mechanism. By continuously monitoring external data sources (such as news, bidding information, and social media), once a new event (such as financing news or technological breakthroughs) is detected, the information extraction process is immediately triggered, and the entities and relationships in the graph are updated in real time.

[0029] Furthermore, an incremental update strategy is adopted during the update process, updating only the changed parts to avoid a full reconstruction and improve efficiency.

[0030] In one embodiment, see Figure 2 The flowchart shown in this embodiment of the present disclosure illustrates a method for constructing a differentiable knowledge graph in the investment and financing field, as follows: Figure 2 As shown, it includes: S201, Define the ontology structure of the investment and financing field, and clarify the set of entity types and relation types in the differentiable knowledge graph.

[0031] In some embodiments, information extraction techniques are used to extract entities and relationships in the investment and financing field from multi-source datasets related to the investment and financing field, and to construct an initial symbolic graph.

[0032] Specifically, it is necessary to define which types of entities and relationships are involved in the investment and financing field.

[0033] For example, entity types include: companies, investors, technologies, industries, products, people, etc.

[0034] Relationship types: investing in, belonging to, competing with, founder, owning technology, upstream and downstream, etc.

[0035] S202 extracts investment and financing facts from multi-source data related to the investment and financing field, forms a set of triplets with (head entity, relationship, tail entity) as units, and constructs an initial symbolic knowledge graph.

[0036] In some embodiments, data extraction and knowledge acquisition are performed. Facts are extracted from multi-source data (business plans, news, financial reports, prospectuses, patent databases, etc.) using natural language processing techniques to form triples of (head entity, relation, tail entity).

[0037] Example: "Company A recently completed its Series B financing, led by Capital C, which will be used for the research and development of its AI medical imaging platform." This can be extracted into the triples (Company A, completed financing, Series B), (Capital C, led investment, Company A), and (Company A, research and development, AI medical imaging platform).

[0038] It should be noted that standardized processing can be performed when extracting data and acquiring knowledge.

[0039] For example, text data (such as business reports and news articles) can be segmented, stop words removed, and entities standardized; while structured data (such as business information) can be mapped to fields and formatted uniformly.

[0040] Next, a symbolic graph is constructed by storing these triples in a graph database (such as Neo4j) or as a list of triples in Simple. At this point, the graph is still symbolic.

[0041] S203, assign a high-dimensional distributed vector representation as the embedding vector for each entity and each relation type in the symbolic knowledge graph.

[0042] In some embodiments, mapping symbols to trainable vectors is key to achieving differentiability, for which a trainable vector is assigned to each unique entity and relation type in the graph.

[0043] For example, creating an embedding matrix: Entity embedding matrix E_entity: Assuming there are 100,000 unique entities, we want to represent each entity using a 128-dimensional vector. Then the entity embedding matrix E_entity is a 100,000x128 matrix. Each row of this matrix represents the initial vector of an entity. This matrix contains trainable parameters.

[0044] The relation embedding matrix E_relation: Assuming there are 50 relation types, also represented by a 128-dimensional vector, then E_relation is a 50x128 matrix. This matrix is ​​also a trainable parameter.

[0045] Initialize the embedding vector: Random initialization: The initial value of a vector can be generated randomly.

[0046] Pre-training initialization: In order to obtain a better starting point, knowledge graph embedding algorithms are usually used to pre-train on existing triplet data.

[0047] That is, knowledge graph embedding technology is used to pre-train on the set of triples of the symbolic knowledge graph to obtain a vector with semantic discriminativeness as the initial value; wherein, the knowledge graph embedding technology is a model based on translation distance, and its pre-training objective function is to minimize the distance between the operation result of the head entity vector, relation vector and tail entity vector in an effective triple and the tail entity vector.

[0048] After pre-training is complete, the resulting vectors are used as the initial values ​​for E_entity and E_relation.

[0049] S204 organizes the embedding vectors of all assigned entities and relations into a trainable parameter matrix and registers this parameter matrix as trainable parameters in the embedded training model.

[0050] In some embodiments, the symbolic knowledge has been parameterized using the above-described scheme.

[0051] The next step is to make it differentiable, including: Build the query interface: Define a function lookup(identifier).

[0052] When an entity ID is input, the function returns the corresponding row vector in the E_entity matrix (such as a 128-dimensional vector).

[0053] When a relation ID is input, the function returns the corresponding row vector in the E_relation matrix.

[0054] Integrating into a deep learning framework: Register the two matrices E_entity and E_relation as parameters embedded in the training model. The embedded training model can be one or more of TransE, DistMult, ComplEx, or RotatE.

[0055] At this point, the knowledge graph is no longer external data, but has become a trainable weight within the model.

[0056] S204, through the vector query interface, the embedding training model can retrieve the corresponding trainable embedding vectors from the parameter matrix by accessing the unique identifiers of entities and relations, and use these vectors as model inputs to participate in forward calculation and backpropagation gradient update, thereby obtaining the differentiable knowledge graph of the investment and financing field.

[0057] In some embodiments, this occurs when the entire system (knowledge graph, graph neural network, and prediction layer) undergoes forward and backward propagation.

[0058] Forward propagation: When performing message passing, the Graph Neural Network (GNN) calls the lookup function to obtain the vectors of nodes (entities) and edges (relationships) for calculation.

[0059] Loss calculation: Calculate the loss of the recommendation results (such as ranking loss).

[0060] Backpropagation: The gradient generated by the loss function will propagate back through the computation graph of the GNN to every element in the E_entity and E_relation matrices.

[0061] Parameter update: Optimizers (such as Adam) update these embedding vectors based on the gradient.

[0062] After optimization, the vector space will change.

[0063] For example, the system might learn: In investment decisions, the concepts of technological barriers and market size should be very close in magnitude.

[0064] The vector of investment relationships should be able to encode the preference patterns of investment institutions very well.

[0065] This means that a knowledge graph is no longer a static database, but an intelligent knowledge base that can adaptively evolve according to specific task objectives (such as precise recommendations).

[0066] Thus, by symbolizing triples, obtaining initial vectors through pre-training, registering vector matrices as model parameters, and performing end-to-end fine-tuning through gradient descent, knowledge graphs are elevated from cold fact bases to components in the model that can learn and optimize themselves, which is the cornerstone for achieving deep and intelligent reasoning.

[0067] In some embodiments, when new triples are extracted from multi-source data, embedding vectors are dynamically assigned to the new entities and relations and added to the trainable parameter matrix, while the gradient calculation and optimization range of the parameter matrix is ​​extended to these new parameters.

[0068] Thus, by parameterizing the symbolic knowledge graph through embedding learning techniques, the semantic information of entities and relations is encoded into optimizable distributed vector representations. This step transforms the discrete symbol space into a continuous vector space, turning the knowledge graph from a static database into a differentiable parametric component within the model. It provides a numerical, semantically rich input foundation for subsequent graph neural network inference and enables the knowledge itself to be dynamically adjusted and refined according to the goals of downstream recommendation tasks.

[0069] S102, semantically align the profile information of the target investment institution with the differentiable knowledge graph, determine the corresponding node of the profile information in the graph, and use it as the starting node for graph reasoning.

[0070] In some embodiments, semantically aligning the profile information of the target investment institution with a differentiable knowledge graph is a crucial first step in initiating the entire graph reasoning process. By finding nodes for each target investment institution, effective information retrieval for that institution is achieved.

[0071] See Figure 3 The flowchart of a semantic alignment method in an embodiment of this disclosure is shown below. Figure 3 As shown, steps S301 to S303 can be performed: S301, Receive profile information of the target investment institution, the profile information including institution name text and / or preference description text.

[0072] S302, the profile information is semantically aligned with the differentiable knowledge graph. Based on the completeness of the profile information, when the institution name in the profile information can uniquely correspond to an existing entity in the differentiable knowledge graph, entity linking technology is used to map the name of the target investment institution to the corresponding specific entity node in the differentiable knowledge graph. When the profile information cannot explicitly correspond to an existing entity in the differentiable knowledge graph, dynamic node generation is performed. A virtual entity node is synthesized in the embedding space of the differentiable knowledge graph using semantic encoding and vector fusion technology.

[0073] In some embodiments, when the target investment institution is an existing entity in the graph, or its profile description is highly precise, a precise semantic alignment path is employed. The core idea is to associate the references in the text description with unique, existing entity nodes in the graph.

[0074] Entity Linking Input: The name or description of the institution provided by the user (e.g., Hillhouse Capital or this top-tier fund headquartered in Beijing and focused on the TMT sector).

[0075] process: Named Entity Recognition (NER): First, extract the key entity names (such as Hillhouse Capital, Beijing, TMT) from the description.

[0076] Candidate entity generation: Retrieve candidate entities with names similar to those extracted from the entity database of the knowledge graph. For example, query entities whose names contain "Gaoling".

[0077] Entity disambiguation: This is the most crucial step. The best-matching entity is selected by calculating the semantic similarity between the text description and the candidate entity and its surrounding information (neighboring nodes).

[0078] For example, the TMT sector mentioned in the description is highly correlated with the information technology and internet sectors connected to the Hillhouse Capital node in the graph, thus confirming the match.

[0079] Output: A uniquely identified graph entity node ID.

[0080] Node localization uses the entity node ID obtained in the previous step as the starting node for graph inference, and loads the current trainable embedding vector of that node for subsequent graph neural network computation.

[0081] In other words, entity linking technology refers to performing named entity recognition on the target investment institution's name text to extract the core institution reference; retrieving a set of candidate entities similar to the core institution reference in the entity library of the differentiable knowledge graph; calculating the semantic similarity between the preference description text and each candidate entity and its neighborhood context, and selecting the candidate entity with the highest semantic similarity as the unique corresponding specific entity node.

[0082] Path 2: Dynamic node generation (suitable for organizations with ambiguous information or entirely new structures) This approach is used when the investment institution is entirely new, or when its profile is relatively abstract (such as a family office that prefers green technology). The core idea is that since there are no directly corresponding nodes in the graph, a temporary virtual node representing its preferences needs to be synthesized in the graph's vector space based on its semantic description.

[0083] Specific implementation steps: Semantic Encoding Input: User-provided description text of the organization's profile.

[0084] Process: A pre-trained language model (such as Sentence-BERT) is used to encode the text into a fixed-dimensional semantic vector, called a profile vector. This vector captures the overall semantics describing the text.

[0085] Output: Profile vector V_profile.

[0086] Nearest Neighbor Search & Vector Fusion: In the embedding space of the knowledge graph, calculate the cosine similarity between V_profile and the vectors of all entity nodes.

[0087] Select the top-K nodes with the highest similarity (these nodes may be related to the same track, technology, region, or existing investment institutions).

[0088] The vectors of these Top-K nodes are weighted and averaged (with similarity scores as the weights), and then fused to generate a new preference vector V_preference. This vector incorporates the existing knowledge in the graph that is most relevant to the description.

[0089] Output: Preference vector V_preference.

[0090] Create a virtual start node: The profile vector V_profile and the preference vector V_preference are fused together using a small neural network to generate the final virtual node vector V_virtual.

[0091] This V_virtual node is added to the current inference session as a temporary node. This virtual node is also considered a trainable parameter and can be fine-tuned in subsequent end-to-end optimizations.

[0092] Output: Virtual starting node and its vector representation.

[0093] In other words, the specific implementation of path two is: The preference description text is encoded into a first semantic vector, denoted as profile vector V_profile, using a pre-trained language model. The semantic similarity between profile vector V_profile and all entity node vectors in the differentiable knowledge graph is calculated, and the top K nodes with the highest similarity are selected as the relevant nearest neighbor set. The vectors of each node in the relevant nearest neighbor set are weighted and summed according to their similarity with V_profile to obtain a second semantic vector, denoted as preference vector V_preference. Profile vector V_profile and preference vector V_preference are concatenated or fused through a fully connected layer to generate a virtual node vector V_virtual representing the target investment institution, and this vector is added to the current computation graph as a trainable parameter.

[0094] S303, the specific entity node or the virtual entity node is determined as the starting node of the graph reasoning.

[0095] For details regarding S303, please refer to the above content.

[0096] S103, taking the starting node as the origin, a graph neural network model is used to perform multiple rounds of message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field. The message passing process uses the embedding vector of the relation type to calculate the importance weight of the neighbor node information, and generates the semantic representation vector of each candidate project node after multiple rounds of iteration.

[0097] In some embodiments, after determining the starting node, multiple iterations are performed with the starting node as the origin to generate a semantic representation vector.

[0098] For example, suppose the starting node is an investment institution X and the candidate project node is a startup P. The goal is to calculate the semantic relevance of P to X.

[0099] In one embodiment, step S103 may include: Obtain the initial feature vectors, relation type embedding vectors, and the starting node for all nodes in the differentiable knowledge graph.

[0100] In some embodiments, each node (including investment institution X and startup P) has its own initial feature vector, such as basic information encoding like the company's industry and stage. Let h_u^k be the representation vector of node u at layer k. Initially, k=0.

[0101] An iterative process of multi-round message passing and information aggregation is executed. In each iteration, for each current node in the differentiable knowledge graph, the following steps are performed: For each neighbor node of the current node, based on the relationship type connecting the neighbor node and the current node, a corresponding relationship embedding vector is selected, and the feature vector of the neighbor node is linearly transformed or nonlinearly mapped using the relationship embedding vector to generate a directed message from the neighbor node to the current node; Based on the feature vector of the current node, the feature vectors of the neighbor nodes, and the relationship embedding vector connecting the two, the attention weight of the directed message is calculated through an attention mechanism; The directed messages from all neighbor nodes are weighted and summed according to their corresponding attention weights to obtain the aggregated neighbor message of the current node; The feature vector of the current node itself is combined with the aggregated neighbor message of the current node, and a new feature vector of the current node in the next iteration is generated through an update function.

[0102] After multiple rounds of preset iterations, the final updated feature vector of each node is output as its semantic representation vector.

[0103] In some embodiments, for each node in the graph (such as startup P), the following operations are performed: Collect messages from neighboring nodes: For each neighboring node v of node P (e.g., the technology field T connected to P, competitor Q, founder F, etc.), prepare a message to be sent from v to P.

[0104] Key implementation (explicitly utilizing relation vectors): This message is not a simple feature of v, but rather modulated by the relation vector r.

[0105] Specifically, the message function is typically: m_{v->u}=W_r*h_v^k, where W_r is a trainable weight matrix corresponding to the relation type r (e.g., belonging to, competing with). This matrix can be viewed as the embodiment of the embedding vector of that relation type in a linear transformation. This means that the transformation of information in the belonging relation and the competing with relation are drastically different.

[0106] Calculate the importance weights of neighbor messages: Not all neighbor messages are equally important. We need to calculate an attention weight a_{v->u}.

[0107] These weights are learned through an attention mechanism, whose inputs include the source node u, the target node v, and the relationship vector e_r between them: a_{v->u} = Attention(h_u^k, e_r, h_v^k). Here, e_r is the embedding vector of the relationship r. This allows the GNN to learn, for example, that investment institution X values ​​information from relationships related to invested projects more than information from geographical location relationships.

[0108] Aggregate message: The messages from all neighbors are summed according to the calculated attention weights to obtain the aggregated message m_u.

[0109] The node state is updated by combining the node's previous layer representation h_u^k with the aggregated neighbor messages m_u. This is done using an update function (such as a fully connected neural network) to generate a new representation h_u^{k+1} for the node in the current layer. This new representation h_u^{k+1} integrates the node's own information and the contextual information from its first-order neighbors.

[0110] Multiple iterations, that is, performing the above steps multiple times, after multiple iterations, the final representation vector h_p^K of each node (such as startup P) contains rich information within its K-hop neighbor range.

[0111] At this point, the vector of startup P has already incorporated the indirect preference information transmitted by investment institution X through the graph structure.

[0112] Finally, the final representation h_x^K of the starting node investment institution X is directly compared with the final representation h_p^K of the candidate startup P. Typically, a vector dot product or concatenation is used to calculate their final matching score via a neural network.

[0113] In this way, by explicitly utilizing the embedding vectors of relation types to calculate the importance weights of neighbor node information, it is possible to distinguish the different contributions of different relation types (such as investing in, belonging to, competing for) to information propagation during message passing, thereby achieving relation-aware semantic aggregation.

[0114] This enables the model to understand and quantify the semantic differences between various relationships in the graph. For example, the model can learn that the relationship of a technology sector favored by investment institutions is far more important than the geographical region where a project is located. Compared to graph algorithms that treat all connections equally, the reasoning mechanism in this embodiment can more accurately capture the deep logical connections that are crucial to investment and financing decisions, thereby generating more insightful semantic representations of projects and laying a solid foundation for high-quality recommendations.

[0115] In one embodiment, the present invention employs a relation-aware graph neural network model to model and analyze the knowledge graph in the investment and financing field.

[0116] A graph neural network model includes at least a node embedding module, a relation type embedding module, a multi-layer message passing and neighborhood aggregation module, and an output prediction module, with each module connected sequentially.

[0117] The message passing and neighborhood aggregation module adopts a multi-layer stacked structure, with the node representation output by the previous layer serving as the input of the next layer between adjacent layers, in order to achieve multi-round message passing.

[0118] The node embedding module is used to vectorize entity nodes in the knowledge graph of the investment and financing field.

[0119] Entity nodes include at least one or more of the following: corporate entities, investment institution entities, natural person entities, industry entities, and financing event entities. For each entity node, the node embedding module generates a corresponding node embedding vector based on the entity's initial features, which include at least: the entity's category identifier; structured attribute information; and a vector representation of the encoded text description features.

[0120] The relation type embedding module is used to vectorize different relation types in the knowledge graph. Relationship types include at least investment and financing relationships such as investment, invested in, employment, affiliation, peer, and holding, and each relation type corresponds to a relation type embedding vector.

[0121] Both node embedding vectors and relation type embedding vectors are used as trainable parameters of the model and are updated through backpropagation during model training.

[0122] The message passing and neighborhood aggregation module includes at least two message passing layers, and the input of each message passing layer is the node embedding vector output by the previous layer.

[0123] In each message passing layer, for the target entity node, the model performs message passing based on the following connection relationship: the target entity node is connected to its neighboring entity nodes through relationship edges; each relationship edge corresponds to a relationship type embedding vector.

[0124] During message passing, the model calculates the influence weight of neighboring entity nodes on the target entity node based on the embedding vector of the target entity node, the embedding vector of neighboring entity nodes, and the embedding vector of the relationship type connecting the two.

[0125] The influence weight is used to reflect the importance of different neighboring entity nodes to the update of the target entity node representation under the current relation semantics, and is constrained by normalization operation.

[0126] The model generates neighborhood message representations based on the embedding vectors of neighborhood entity nodes and the corresponding relationship type embedding vectors, and performs weighted aggregation on multiple neighborhood messages according to the influence weights to obtain the neighborhood aggregation result of the target entity node.

[0127] Subsequently, the neighborhood aggregation result is fused with the current embedding vector of the target entity node to generate an updated representation of the target entity node, which serves as the input to the next message passing layer.

[0128] Through multi-layer message passing and neighborhood aggregation, the model integrates investment and financing association information of target entity nodes within a multi-hop neighborhood range layer by layer.

[0129] After the graph neural network model is constructed, it is trained based on historical data in the investment and financing field.

[0130] In one implementation, the training data is constructed in the form of investment and financing relationship triples, including positive sample relationships and negative sample relationships constructed through negative sampling.

[0131] During model training, the final embedding vectors of entity nodes and relation type embedding vectors are used as inputs to calculate the prediction results, and a loss function is constructed based on the difference between the prediction results and the true annotations.

[0132] The model uses gradient descent to jointly optimize node embedding vectors, relation type embedding vectors, and network parameters at each layer.

[0133] The key parameters of the model include at least: 2 to 6 message passing layers; 64 to 512 node embedding vector dimensions; 32 to 256 relation type embedding vector dimensions; and regularization and random deactivation mechanisms can be introduced during training to prevent overfitting.

[0134] S104, calculate the relevance score between the candidate items and the starting node based on the semantic representation vector to generate a ranking list, and decode the key path activated during the inference process of the graph neural network model as interpretability context, input it into the domain expert model, and generate corresponding recommendation reasons and scores.

[0135] In some embodiments, this is a key step in transforming black-box reasoning from graph neural networks into interpretable and trustworthy recommendation results.

[0136] Specifically, the above method yields the final semantic representation vector representing the starting node (i.e., the target investment institution) and the final semantic representation vector representing the i-th candidate project node. This allows us to determine the corresponding recommendation reasons and scores.

[0137] In some embodiments, the domain expert model is built on a natural language model, which employs a neural network structure that includes at least an embedding layer, an encoding layer, and an output layer.

[0138] An embedding layer is used to map input text or structured context information into a continuous vector representation. The embedding layer includes at least word embeddings, position embeddings, and optional paragraph or role embeddings.

[0139] The encoding layer includes a multi-layer attention mechanism network structure, which is used to model the semantic information in the input context and capture the relationships between different contextual elements.

[0140] The output layer is used to generate model output based on the contextual representation output by the encoding layer, including the recommendation reason text in natural language form and the corresponding rating information.

[0141] The aforementioned embedding layer, encoding layer, and output layer are connected in sequence, with the output of the previous layer serving as the input of the next layer.

[0142] The embedding layer includes at least the following sub-embedding units: A lexical embedding unit is used to map lexical units in the input text into word vectors of fixed dimensions. The lexical units include natural language vocabulary, entity identifiers, and relation type identifiers.

[0143] The position embedding unit is used to introduce positional information into each word in the input sequence to represent the sequential relationship of words in the key path description or context text.

[0144] Structural role embedding units are used to distinguish different types of input information.

[0145] In one implementation, different structural role embedding vectors are assigned to entity nodes, relation types, weight identifiers, and prompt instructions in the critical path. The word embeddings, position embeddings, and structural role embeddings are fused by vector addition or concatenation to form the output representation of the embedding layer.

[0146] The dimension of the embedding vector is preferably set between 256 and 2048 to balance expressive power and computational efficiency.

[0147] The encoding layer consists of a multi-layered attention network structure, with each layer connected sequentially from bottom to top.

[0148] Each coding layer includes at least: Multi-head self-attention sublayer is used to calculate the relevance weights between different tokens in the input sequence, thereby capturing the dependencies between entities, relations and weights in the critical path.

[0149] Different attention heads are used to focus on different types of semantic relationships, such as causal associations between entities, relationship chains, or risk constraint information.

[0150] The feedforward neural network sublayer is used to perform nonlinear transformations on the output of the attention sublayer to enhance the model's ability to express complex investment and financing logic.

[0151] The residual connection and normalization unit is used to fuse sublayer inputs and outputs and perform normalization operations to improve the stability of model training and inference.

[0152] The number of coding layers is preferably set to 6 to 48, with each layer sharing the same basic structure but with independent parameters.

[0153] The output layer includes at least: The scoring output unit is used to generate a numerical scoring result for investment and financing recommendations based on the context vector output by the encoding layer. The scoring result can be a continuous value or a discrete level, used to represent the matching degree or credibility of the investment and financing recommendations.

[0154] The reason generation unit is used to generate natural language reasons for recommendation based on the contextual semantic representation output by the encoding layer.

[0155] The reason generation unit outputs text through sequence generation and focuses on the entity and relationship information involved in the critical path during the generation process.

[0156] In one implementation, the scoring output unit and the reason generation unit share the semantic representation of the coding layer output, but use different output header structures.

[0157] In some embodiments, step S104 may include: Calculate the similarity between the semantic representation vector of the starting node and the semantic representation vector of each candidate item node to obtain the relevance score of each candidate item, and sort all candidate items in descending order based on the score to generate a sorted list of items.

[0158] This step is relatively straightforward; the goal is to obtain a quantified matching degree. Specific implementation steps: After multiple rounds of message passing, two key feature vectors were obtained: h_start: Represents the final semantic representation vector of the starting node (i.e., the target investment institution).

[0159] h_candidate_i: Represents the final semantic representation vector of the i-th candidate item node.

[0160] Calculate the similarity between two vectors, such as cosine similarity, dot product, or concatenate or subtract the two vectors and input them into a small fully connected network to output a score.

[0161] In this way, a relevance score score_i can be calculated for each candidate item i.

[0162] Sort all candidate projects in descending order of their score_i, and select the top-N projects to form the final project recommendation ranking list.

[0163] For candidate projects ranked above a predetermined threshold in the project ranking list, the information propagation path with the highest attention weight in the multi-round message passing process of the graph neural network model is traced back, and the path sequence composed of nodes and relationships is extracted.

[0164] In graph neural networks, the attention mechanism assigns a weight to the messages from each neighboring node during message passing. This weight intuitively reflects which neighboring node's information is more important when forming the current node's perception.

[0165] For each candidate item that ranks highly in the sort, backtrack backward from its node h_candidate_i back to the starting node h_start.

[0166] At each step of the backtracking, the edge with the highest attention weight is selected (i.e., which relationship or which neighbor node contributes the most to the current node).

[0167] For example, to understand why project P scored highly, backtracking revealed that GNN is most concerned with the edge that project P belongs to the AI ​​track, and the information of the AI ​​track nodes mainly comes from the edge that the starting institution focuses on the AI ​​track.

[0168] The node and relation names in the extracted path sequence are linearized into natural language descriptions according to their logical order in the path, forming a structured interpretable context.

[0169] The backtracked node and relation sequence (a subgraph structure) is converted into a natural language description that can be understood by a large language model.

[0170] For example, concatenating the node names and relation names on a path into a coherent sentence.

[0171] The path obtained by backtracking is: [Starting node: Institution X] — (Focus on) — -> [Node A: AI track] <-- (Belongs to) — [Candidate node: Project P].

[0172] Linearized description: Generated text: Institution X focuses on the AI ​​track, while project P belongs to the AI ​​track.

[0173] The interpretability context, basic information of candidate projects, and task instructions are combined to construct prompt words, which are then input into the domain expert model to generate recommendation reasons and scores for each candidate project. The goal is to guide natural language models to generate professional and accurate recommendation reasons.

[0174] The linearized path description is used as the core context and embedded into a carefully designed prompt word template.

[0175] For example, the prompt word: You are an investment and financing recommendation expert. Please generate a concise recommendation for an investment institution based on the following reasoning path.

[0176] Investment Institution Profile: {Institution Name}, Areas of Interest: {List of Areas}, Invested Projects: {List of Projects}.

[0177] Candidate project information: {Project Name}, Technical Highlights: {Technology Description}, Industry: {Industry}, Related Entities: {Companies, Technologies, Competitors, etc. associated in the graph}.

[0178] Please assess the compatibility of this project with the investment institution and provide reasons for your recommendation.

[0179] Optionally, multi-hop relationship subgraphs linking projects can be extracted from the knowledge graph and embedded as textual prompts. For example, external information such as industry growth rates, policy updates, and competitive landscape can be incorporated.

[0180] The constructed prompts are input into the domain expert model to generate fluent and natural recommendation reasons and ratings.

[0181] Example output: The main reason for recommending this project is its high alignment with your organization's strategic layout. Your organization has long focused on the artificial intelligence sector, and this project's core business falls within this field, possessing significant potential for technological synergy and market development prospects.

[0182] In one example, the output of the domain expert model can be in JSON format, such as: json { "Match score": 0.92 Reasons for recommendation: [Reason 1, Reason 2, Reason 3] }

[0183] In some embodiments, project information, matching degree, and recommendation reasons are packaged and returned to the front end to support users in clicking on entities in the reasons to view source information in the knowledge graph.

[0184] In some embodiments, a minimum threshold (such as 0.7) can be set to filter low-scoring items.

[0185] In some embodiments, result filtering and deduplication can be performed. For example, projects that a user has already invested in or rejected can be removed; or duplicate projects (such as different funding rounds of the same company) can be merged.

[0186] By decoding the key paths activated during the reasoning process of a graph neural network and using them as interpretable context input to a domain expert model to generate recommendation reasons, this method combines the model's internal, symbolic reasoning chain with external natural language generation capabilities. It provides transparent and credible generation justification for each recommendation, allowing users (investment managers) to clearly understand the logic behind their decisions, rather than facing an incomprehensible black box recommendation.

[0187] In other words, this embodiment not only provides the answer to what to recommend, but more importantly, it systematically answers the key question of why to recommend, greatly enhancing the credibility and practicality of the recommendation system and helping users make more informed decisions.

[0188] In some embodiments, the domain expert model is obtained by: collecting text data in the investment and financing field and constructing a fine-tuning sample set of instruction-answer pairs; using low-rank adaptive technology, based on the fine-tuning sample set, fine-tuning the natural language model according to instructions with the goal of maximizing accuracy and logical consistency in investment and financing related tasks, to obtain a domain expert model for investment and financing.

[0189] More specifically, textual data in the investment and financing field is collected, including question-and-answer pairs, project analysis reports, and investment logic documents; based on the textual data, a high-quality fine-tuned sample set containing at least 100,000 instruction-answer pairs is constructed through manual annotation or large model synthesis.

[0190] Efficient parameter fine-tuning of the basic large language model: Using LoRA (low-rank adaptive) technology, with the training objective of maximizing accuracy and logical consistency in investment and financing related tasks, the LLaMA-7B model is fine-tuned to obtain an expert model in the investment and financing field.

[0191] Furthermore, the fine-tuned investment and financing expert model is evaluated using a reserved test set, with the evaluation tasks including at least project matching and investment rationale generation; the qualified model is deployed as an API service for external recommendation modules to call.

[0192] In some embodiments, LoRA technology is used for fine-tuning, specifically including: injecting a trainable low-rank decomposition matrix into the attention mechanism module of the LLaMA-7B model, freezing the original parameters of the model, and training only the low-rank decomposition matrix.

[0193] During fine-tuning, the cross-entropy loss function is used, and a logical consistency reward signal is introduced. The model output is optimized through reinforcement learning or weighted loss.

[0194] The evaluation specifically includes a combination of automatic and manual evaluation, with automatic evaluation metrics including at least BLEU, ROUGE, and accuracy scores based on domain expert annotations.

[0195] In some embodiments, considering the existence of erroneous outputs, the investment and financing project recommendation method based on dynamic knowledge graphs and large language models may further include: obtaining user interaction feedback on the project ranking list and recommendation reasons, and converting it into a supervision signal; using the gradient backpropagation algorithm, optimizing the embedding vector of the differentiable knowledge graph, the parameters of the graph neural network model, and the generation strategy of the domain expert model using the supervision signal, and outputting the optimized final project ranking list and recommendation reasons.

[0196] By employing a loop optimization mechanism—which collects user interaction feedback, constructs a joint loss function, and utilizes gradient backpropagation to simultaneously optimize the parameters of the knowledge graph, graph neural network, and large language model—the system can continuously learn from each human-computer interaction. This allows the three core components to co-evolve around a unified optimization goal (i.e., satisfying user preferences): the semantic representation of the knowledge graph becomes more aligned with domain needs based on feedback, the reasoning logic of the graph neural network becomes more accurate, and the interpretation of the large language model becomes more precise. In this way, a self-driven, continuously iteratively evolving agent is realized, capable of adapting to long-term changes in user preferences and market trends, providing increasingly accurate and personalized services.

[0197] In some embodiments, the feedback correction method may include: Collect user interaction feedback data with the system's recommendation results. The feedback data includes, but is not limited to, actions such as clicking, ignoring, and deep browsing of recommended items, as well as actions such as liking, disliking, and modifying the reasons for generating recommendations.

[0198] The interactive feedback data is transformed into a multi-task joint loss function, which includes at least a ranking loss term based on the item ranking list and an explanation loss term based on the quality of the recommendation reasons.

[0199] Among them, the joint loss function ; in, This is a list-level sorting loss used to penalize item sorting that is inconsistent with user feedback; This is used to penalize recommendations that receive negative user reviews, citing quality loss as the reason. This is a regularization term used to prevent overfitting; These are hyperparameters used to balance the loss weights.

[0200] in, A ranking loss function based on sample pairs or sample lists is adopted, and its input is the item relevance score calculated by the graph neural network and the real ranking label composed of user feedback.

[0201] In the deep learning framework, the differentiable gradient of the joint loss function with respect to the entity and relation embedding vectors in the differentiable knowledge graph, the weights of the graph neural network model, and the parameters of the domain expert model generation layer is calculated.

[0202] Using the gradient descent algorithm, the embedding parameters of the differentiable knowledge graph, the model weights of the graph neural network, and the generation strategy parameters of the large language model are updated simultaneously based on the calculated gradient.

[0203] Repeat the above steps to ensure that the output item ranking list and recommendation reasons continuously approximate the user's true preferences and decision-making logic in subsequent iterations.

[0204] In summary, this embodiment of the disclosure integrates a full-chain technology, including knowledge-differentiable representation, relation-aware reasoning, path-interpretable generation, and feedback closed-loop optimization, to construct an end-to-end investment and financing recommendation system that ranges from multi-source information perception to intelligent decision generation and from static knowledge utilization to dynamic model evolution.

[0205] This method upgrades traditional recommendation schemes that rely on keyword matching, static rules, or shallow collaborative filtering to a semantic reasoning system based on dynamic knowledge graphs and deep graph neural networks, forming an intelligent recommendation mechanism that can adapt to investment institution preferences, market dynamics, and project characteristics.

[0206] By introducing differentiable knowledge graphs, the system transforms symbolic domain knowledge into optimizable model parameters, providing a semantically rich and computable foundation for deep reasoning. Through relation-aware graph neural networks, it achieves deep modeling of investment logic and mining of complex relationships. By fusing key path decoding with large language models, it achieves credible traceability and natural language interpretation of recommendation results. Finally, through user feedback-driven joint loss functions and end-to-end gradient optimization, it achieves the collaborative self-evolution of system parameters.

[0207] This system fundamentally solves the bottlenecks of traditional solutions in terms of semantic understanding depth, decision logic transparency, and system adaptability, thereby significantly improving the accuracy, interpretability, and effectiveness of decision support in highly complex and unstructured investment and financing decision-making environments.

[0208] The above describes in detail the investment and financing project recommendation method based on dynamic knowledge graph and large language model through some embodiments. In order to enable those skilled in the art to better understand and implement it, the corresponding system is also described in detail below through some embodiments.

[0209] See Figure 4 The diagram shown below illustrates the structure of an investment and financing project recommendation system based on dynamic knowledge graphs and large language models in an embodiment of this disclosure. Figure 4 As shown, the investment and financing project recommendation system 400 based on dynamic knowledge graphs and large language models may include: Construction unit 410 is used to construct a differentiable knowledge graph in the field of investment and financing, wherein each entity node and relation type in the differentiable knowledge graph is parameterized as an embedding vector. Alignment unit 420 is used to semantically align the profile information of the target investment institution with the differentiable knowledge graph, determine the corresponding node of the profile information in the graph, and use it as the starting node for graph reasoning. The iteration unit 430 is used to perform multiple rounds of message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field with the starting node as the origin and a graph neural network model. The message passing process uses the embedding vector of the relation type to calculate the importance weight of the neighbor node information and generates the semantic representation vector of each candidate project node after multiple rounds of iteration. Recommendation unit 440 is used to calculate the relevance score between candidate items and the starting node based on the semantic representation vector to generate a ranking list, and decode the key path activated during the graph neural network inference process as interpretability context, input it into the domain expert model, and generate corresponding recommendation reasons and scores.

[0210] For further details regarding the construction unit 410, alignment unit 420, iteration unit 430, and recommendation unit 440, please refer to the aforementioned examples.

[0211] It is understandable that the above division of units is only a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, the above units can be implemented by the processor calling software.

[0212] This disclosure also provides an electronic device for implementing an investment and financing project recommendation method based on dynamic knowledge graphs and large language models.

[0213] The electronic device includes a memory and a processor, as well as a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any of the foregoing descriptions.

[0214] It should be noted that the computer system of the electronic device shown below is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0215] like Figure 5The schematic diagram shown in this embodiment illustrates the hardware structure of an electronic device 500. The electronic device 500 includes a Central Processing Unit (CPU) 501, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 502 or loaded from storage portion 508 into Random Access Memory (RAM) 503, such as executing the methods described in the above embodiments. The RAM 503 also stores various programs and data required for system operation. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0216] The following components are connected to I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. Drive 55 is also connected to I / O interface 505 as needed. Removable media 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 55 as needed so that computer programs read from them can be installed into storage section 508 as needed.

[0217] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs various functions defined in the system of this application.

[0218] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0219] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations described above.

[0220] In another aspect, this application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

[0221] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0222] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0223] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0224] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0225] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0226] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for recommending investment and financing projects based on dynamic knowledge graphs and large language models, characterized in that, include: Construct a differentiable knowledge graph in the field of investment and financing, wherein each entity node and relation type in the differentiable knowledge graph is parameterized as an embedding vector; Semantically align the profile information of the target investment institution with the differentiable knowledge graph to determine the corresponding node of the profile information in the graph, and use it as the starting node for graph reasoning; Using the starting node as the origin, a graph neural network model is used to perform multiple rounds of message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field. The message passing process uses the embedding vector of the relation type to calculate the importance weight of the neighbor node information, and generates the semantic representation vector of each candidate project node after multiple rounds of iteration. Based on the semantic representation vector, the relevance score between the candidate items and the starting node is calculated to generate a ranking list. The key path activated during the graph neural network model is decoded as interpretability context and input into the domain expert model to generate corresponding recommendation reasons and scores.

2. The method for recommending investment and financing projects according to claim 1, characterized in that, The construction of a differentiable knowledge graph in the investment and financing field includes: Define the ontology structure of the investment and financing field, and determine the set of entity types and relation types in the differentiable knowledge graph; Investment and financing facts are extracted from multi-source data related to the investment and financing field, forming a set of triplets with (head entity, relationship, tail entity) as the unit, and an initial symbolic knowledge graph is constructed. For each entity and each relation type in the symbolic knowledge graph, a high-dimensional distributed vector representation is assigned as the embedding vector; The embedding vectors of all assigned entities and relations are organized together into a trainable parameter matrix, and this parameter matrix is ​​registered as trainable parameters in the embedded training model. Through the vector query interface, the embedding training model can retrieve the corresponding trainable embedding vectors from the parameter matrix by accessing the unique identifiers of entities and relations, and use these vectors as model inputs to participate in forward computation and backpropagation gradient updates, thereby obtaining the differentiable knowledge graph of the investment and financing field.

3. The method for recommending investment and financing projects according to claim 2, characterized in that, Meet one or more of the following conditions: Knowledge graph embedding technology is used to pre-train on the set of triples of the symbolic knowledge graph to obtain a semantically discriminative vector as the initial value; wherein, the knowledge graph embedding technology is a translation distance-based model, and its pre-training objective function is to minimize the distance between the operation result of the head entity vector, relation vector and tail entity vector in an effective triple and the tail entity vector. When new triples are extracted from multi-source data, embedding vectors are dynamically assigned to the new entities and relations and added to the trainable parameter matrix. At the same time, the gradient calculation and optimization range of the parameter matrix is ​​extended to these new parameters.

4. The method for recommending investment and financing projects according to claim 1, characterized in that, The step of semantically aligning the profile information of the target investment institution with the differentiable knowledge graph, determining the corresponding node of the profile information in the graph, and using it as the starting node for graph reasoning includes: Receive profile information of the target investment institution, the profile information including the institution's name text and / or preference description text; The profile information is semantically aligned with the differentiable knowledge graph. Based on the completeness of the profile information, when the institution name in the profile information can uniquely correspond to an existing entity in the differentiable knowledge graph, entity linking technology is used to map the name of the target investment institution to the corresponding specific entity node in the differentiable knowledge graph. When the profile information cannot explicitly correspond to an existing entity in the differentiable knowledge graph, dynamic node generation is performed. A virtual entity node is synthesized in the embedding space of the differentiable knowledge graph using semantic encoding and vector fusion technology. The specific entity node or the virtual entity node is determined as the starting node of the graph reasoning.

5. The method for recommending investment and financing projects according to claim 1, characterized in that, The process of using the starting node as the origin and employing a graph neural network model to perform multi-round message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field includes: Obtain the initial feature vectors, relation type embedding vectors, and the starting node for all nodes in the differentiable knowledge graph; An iterative process of multi-round message passing and information aggregation is executed. In each iteration, for each current node in the differentiable knowledge graph, the following steps are performed: For each neighbor node of the current node, based on the relationship type connecting the neighbor node and the current node, a corresponding relationship embedding vector is selected, and the feature vector of the neighbor node is linearly transformed or nonlinearly mapped using the relationship embedding vector to generate a directed message from the neighbor node to the current node; Based on the feature vector of the current node, the feature vectors of the neighbor nodes, and the relationship embedding vector connecting the two, the attention weight of the directed message is calculated through an attention mechanism; The directed messages from all neighbor nodes are weighted and summed according to their corresponding attention weights to obtain the aggregated neighbor message of the current node; The feature vector of the current node itself is combined with the aggregated neighbor message of the current node, and a new feature vector of the current node in the next iteration is generated through an update function; After multiple rounds of preset iterations, the final updated feature vector of each node is output as its semantic representation vector.

6. The method for recommending investment and financing projects according to claim 1, characterized in that, The process involves calculating the relevance score between candidate items and the starting node based on the semantic representation vector to generate a ranked list, and decoding the key paths activated during the graph neural network model's inference process as interpretable context. This is then input into the domain expert model to generate corresponding recommendation reasons and scores, including: Calculate the similarity between the semantic representation vector of the starting node and the semantic representation vector of each candidate item node to obtain the relevance score of each candidate item, and sort all candidate items in descending order based on the score to generate a sorted list of items. For candidate projects ranked above a predetermined threshold in the project ranking list, the information propagation path with the highest attention weight in the multi-round message passing process of the graph neural network model is traced back, and the path sequence composed of nodes and relationships is extracted. The node and relation names in the extracted path sequence are linearized into natural language descriptions according to their logical order in the path, forming a structured interpretable context. The interpretability context, basic information of candidate projects, and task instructions are combined to construct prompt words, which are then input into the domain expert model to generate recommendation reasons and scores for each candidate project.

7. The method for recommending investment and financing projects according to claim 6, characterized in that, Also includes: Obtain user feedback on the project ranking list and recommendation reasons, and convert it into monitoring signals; The gradient backpropagation algorithm is used to optimize the embedding vector of the differentiable knowledge graph, the model parameters of the graph neural network, and the generation strategy of the domain expert model using the supervision signal, and outputs the optimized final project ranking list and recommendation reasons.

8. The method for recommending investment and financing projects according to any one of claims 1 to 6, characterized in that, The domain expert model is obtained in the following way: Collect textual data in the investment and financing field and construct a fine-tuned sample set of instruction-answer pairs; Using low-rank adaptive technology, and based on the fine-tuning sample set, with the training objective of maximizing accuracy and logical consistency in investment and financing-related tasks, the natural language model is fine-tuned to obtain a domain expert model for investment and financing.

9. A recommendation system for investment and financing projects based on dynamic knowledge graphs and large language models, characterized in that, include: A construction unit is used to construct a differentiable knowledge graph in the field of investment and financing, wherein each entity node and relation type in the differentiable knowledge graph is parameterized as an embedding vector. The alignment unit is used to semantically align the profile information of the target investment institution with the differentiable knowledge graph, determine the corresponding node of the profile information in the graph, and use it as the starting node for graph reasoning. The iterative unit is used to perform multiple rounds of message passing and neighborhood information aggregation on the knowledge graph of the investment and financing field with the starting node as the origin and a graph neural network model. The message passing process uses the embedding vector of the relation type to calculate the importance weight of the neighbor node information, and generates the semantic representation vector of each candidate project node after multiple rounds of iteration. The recommendation unit is used to calculate the relevance score between the candidate items and the starting node based on the semantic representation vector to generate a ranked list, and decode the key path activated in the process of the graph neural network model as interpretability context, input it into the domain expert model, and generate corresponding recommendation reasons and scores.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor runs the computer program, it performs the steps of the investment and financing project recommendation method based on dynamic knowledge graph and large language model as described in any one of claims 1 to 8.

Images