Transaction hypothesis generation method and device based on semantic diffusion, equipment and storage medium

Abstract

The application provides a transaction hypothesis generation method and device based on semantic diffusion, equipment and storage medium, the method obtains query task information for a target object, the target object is a transaction research target, the query task information indicates query guide information associated with the transaction research target, the target object is taken as a seed node, and the query task information is taken as a semantic guide to perform semantic diffusion propagation query in a financial time heterogeneous graph, to obtain a plurality of target nodes associated with the seed node; the financial time heterogeneous graph includes a plurality of types of nodes and a plurality of edges having a plurality of association relationship types connecting different nodes; each node and each edge has a time label; an evidence subgraph is constructed based on the plurality of target nodes and the edges between the nodes, and message aggregation is performed to obtain an evidence vector of the evidence subgraph, and based on the evidence subgraph and the evidence vector, hypothesis generation processing is performed to obtain a transaction hypothesis. The application embodiment reduces the introduction of irrelevant nodes and can reduce the amount of calculation.

Description

Technical Field

[0001] This application relates to the field of semantic diffusion query technology, and more specifically, to a method, apparatus, device, and storage medium for generating transaction hypotheses based on semantic diffusion. Background Technology

[0002] With the rapid growth of financial information, how to efficiently and accurately retrieve information nodes related to the user's query intent from massive and multi-source financial data, and organize them into a usable chain of evidence, has become an important technical problem in the field of fintech.

[0003] In related technologies, information retrieval methods based on financial knowledge graphs are commonly used. The implementation of these methods relies on the self-attribute information reflected by the graph nodes for querying, which often leads to a large number of irrelevant nodes being included in the query scope, resulting in significant computational overhead during the retrieval process. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, device and storage medium for generating transaction hypotheses based on semantic diffusion, so as to at least solve the problems existing in the related technologies.

[0005] Specifically, this application is implemented through the following technical solution: This application provides a transaction hypothesis generation method based on semantic diffusion, including: Obtain query task information for a target object; the target object is a transaction research target, and the query task information indicates query guidance information associated with the transaction research target; Using the target object as a seed node, semantic diffusion propagation queries are performed in the financial time heterogeneous graph, guided by the query task information, to obtain multiple target nodes associated with the seed node. The financial time heterogeneous graph includes multiple types of nodes and multiple edges connecting different nodes with multiple types of association relationships. Each node and each edge has a time label. The multiple types of nodes include multiple types of object nodes, information nodes, attribute nodes, and decision nodes. The multiple types of association relationships include multiple types of temporal association relationships, hierarchical association relationships, causal association relationships, information association relationships, and logical association relationships. Based on the multiple target nodes and the edges between the nodes, an evidence subgraph is constructed, and message aggregation is performed on the evidence subgraph to obtain the evidence vector of the evidence subgraph; Based on the evidence subgraph and the evidence vector, hypothesis generation processing is performed on the target object to obtain the transaction hypothesis corresponding to the target object.

[0006] This application also provides a transaction hypothesis generation device based on semantic diffusion, comprising: The information acquisition module is used to acquire query task information for a target object; the target object is a transaction research target, and the query task information indicates query guidance information associated with the transaction research target; The semantic query module is used to take the target object as a seed node and perform semantic diffusion and propagation queries in the financial time heterogeneous graph, guided by the query task information, to obtain multiple target nodes associated with the seed node. The financial time heterogeneous graph includes multiple types of nodes and multiple edges with multiple association types connecting different nodes. Each node and each edge has a time label. The multiple types of nodes include multiple types of object nodes, information nodes, attribute nodes, and decision nodes. The multiple association types include multiple types of temporal association, hierarchical association, causal association, information association, and logical association. The message aggregation module is used to construct an evidence subgraph based on the multiple target nodes and the edges between the nodes, and to perform message aggregation on the evidence subgraph to obtain the evidence vector of the evidence subgraph. The hypothesis generation model is used to perform hypothesis generation processing on the target object based on the evidence subgraph and the evidence vector to obtain the transaction hypothesis corresponding to the target object.

[0007] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the semantic diffusion-based transaction hypothesis generation method described in any of the foregoing embodiments.

[0008] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the semantic diffusion-based transaction hypothesis generation method described in any of the foregoing embodiments.

[0009] This application also provides a computer program product, including a computer program that, when run by a processor, performs the steps of any of the possible semantic diffusion-based transaction hypothesis generation methods described above.

[0010] The technical solutions provided by the embodiments of this application may include the following beneficial effects: In this embodiment, semantic diffusion propagation query is performed on a temporally heterogeneous graph by using query task information as semantic guidance. The propagation process uses query task information as semantic guidance, giving the graph traversal process a clear query direction and filtering direction, significantly reducing the introduction of irrelevant nodes, thereby achieving control over the propagation scale, effectively reducing the number of nodes that the computer device needs to traverse and process during propagation, reducing the storage and computing overhead of graph data processing, improving the processing efficiency of the computer system, and overcoming the static defects of fixed hop count search or keyword retrieval in the prior art.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this specification. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a transaction hypothesis generation method based on semantic diffusion, as shown in an exemplary embodiment of this application; Figure 2 This is a flowchart illustrating a semantic diffusion propagation query as shown in an exemplary embodiment of this application; Figure 3 This is a flowchart illustrating another transaction hypothesis generation method based on semantic diffusion, as shown in an exemplary embodiment of this application; Figure 4 This is a schematic diagram of the structure of a transaction hypothesis generation device based on semantic diffusion, as illustrated in an exemplary embodiment of this application. Figure 5 This is a schematic diagram of another transaction hypothesis generation device based on semantic diffusion, as illustrated in an exemplary embodiment of this application. Figure 6 This is a hardware structure diagram of a computer device illustrated in an exemplary embodiment of this application. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0014] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0015] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0016] With the rapid growth of financial information, how to efficiently and accurately retrieve information nodes related to the user's query intent from massive and multi-source financial data, and organize them into a usable chain of evidence, has become an important technical problem in the field of fintech.

[0017] In related technologies, information retrieval methods based on financial knowledge graphs are commonly used. The implementation of these methods relies on the self-attribute information reflected by the graph nodes for querying, which often leads to a large number of irrelevant nodes being included in the query scope, resulting in significant computational overhead during the retrieval process.

[0018] Specifically, information retrieval methods based on financial knowledge graphs typically employ the following technical approach: constructing a knowledge graph containing financial entities and their relationships; receiving query text input by the user; and performing keyword matching or full-text search on the graph database after word segmentation to obtain nodes and paths related to the query. These methods primarily rely on literal keyword matching during information retrieval, resulting in a low semantic relevance between the query results and the user's query intent.

[0019] Search-enhanced generation (RAG) technology typically uses static knowledge bases for information retrieval, directly retrieving knowledge from the entire knowledge base or database. This results in a large number of nodes and knowledge that are irrelevant to the query intent in the search results, reducing retrieval efficiency and the relevance between the search results and the user's query target.

[0020] In the field of knowledge graph querying, existing query methods are mostly based on subgraph matching of node labels. Since node labels only reflect the entity's name or attributes and other self-information, they are insufficient to reflect the deep semantic relationships between nodes, resulting in insufficient semantic relevance of the query results and easily introducing a large number of entity nodes with weak relevance to the query intent. Furthermore, existing knowledge graph reasoning methods typically include a large number of entities in the computational scope during propagation, which can easily lead to high storage and computational overhead and may cause oversmoothing; as the number of propagation steps increases, entities and facts with weak relevance to the current query relationship are easily introduced, increasing evidentiary noise.

[0021] In financial scenarios, these problems are particularly prominent. Financial information is characterized by diverse sources, complex relationships, and high time sensitivity. Users' research intentions often point to specific types of relationships (such as supply chain transmission, peer influence, etc.). However, existing financial knowledge graph retrieval methods, upon receiving a user query, typically use subgraph search with a fixed number of hops or global keyword matching to obtain candidate nodes. This makes it difficult to dynamically adjust the retrieval neighborhood according to different query intentions and relationship types, easily introducing a large number of nodes unrelated to the target relationship into the search results. This dilutes the evidence signal and affects the accuracy and reliability of subsequent analysis.

[0022] In summary, existing technologies have at least the following drawbacks: First, information retrieval relies on static keyword matching or node tag matching, lacking the ability to dynamically understand the semantics of queries; second, the retrieval process is difficult to target based on the type of association in the query intent, easily introducing a large number of irrelevant nodes, resulting in low retrieval accuracy and high noise; third, there is a lack of a mechanism to effectively utilize the temporal information of nodes during the node retrieval process, making it difficult to meet the timeliness requirements of information in financial scenarios.

[0023] Based on the above research, this disclosure provides a transaction hypothesis generation method based on semantic diffusion. The method first obtains query task information for a target object; the target object is the transaction research object, and the query task information indicates query guidance information associated with the transaction research object. Second, using the target object as a seed node, semantic diffusion propagation is performed in a financial time heterogeneous graph using the query task information as semantic guidance to obtain multiple target nodes associated with the seed node. The financial time heterogeneous graph includes multiple types of nodes and multiple edges connecting different nodes with multiple types of association relationships. Each node and each edge has a time label. The multiple types of nodes include multiple types of object nodes, information nodes, attribute nodes, and decision nodes. The multiple types of association relationships include multiple types of temporal association relationships, hierarchical association relationships, causal association relationships, information association relationships, and logical association relationships. Then, based on the multiple target nodes and the edges between nodes, an evidence subgraph is constructed, and message aggregation is performed on the evidence subgraph to obtain the evidence vector of the evidence subgraph. Finally, based on the evidence subgraph and the evidence vector, hypothesis generation processing is performed on the target object to obtain the transaction hypothesis corresponding to the target object.

[0024] In this embodiment, semantic diffusion propagation query is performed on a time-series heterogeneous graph by using query task information as semantic guidance. The propagation process uses query task information as semantic guidance, giving the graph traversal process a clear query direction and filtering direction, significantly reducing the introduction of irrelevant nodes, thereby achieving control over the propagation scale, effectively reducing the number of nodes that the computer system needs to traverse and process during propagation, reducing the storage and computing overhead of graph data processing, improving the processing efficiency of computer equipment, and overcoming the static defects of fixed hop count search or keyword retrieval in the prior art.

[0025] Furthermore, reducing irrelevant nodes can improve the precision and accuracy of evidence discovery.

[0026] Secondly, each node and side of the financial time heterogeneous graph has a time label, which allows the propagation process to be filtered and controlled according to the time dimension.

[0027] Finally, the above method realizes the process of query retrieval, semantic diffusion propagation, evidence subgraph extraction and aggregation, and structured hypothesis output. It can automatically complete the entire process transformation from raw query input to executable decision hypothesis without manual information organization or rule writing, thus bridging the technical gap between graph reasoning and structured output.

[0028] To facilitate understanding of this embodiment, a detailed description of the semantic diffusion-based transaction hypothesis generation method disclosed in this disclosure is provided first. The execution entity of this semantic diffusion-based transaction hypothesis generation method is generally a computer device. This computer device can be a server, which can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data, and artificial intelligence platforms. In other embodiments, the computer device can also be a terminal device, which can be a mobile device, terminal, handheld device, computing device, vehicle-mounted device, etc.

[0029] In other embodiments, the method can also be applied to an implementation environment consisting of computer equipment and servers, or an implementation environment consisting of terminal equipment and servers. Furthermore, this semantic diffusion-based transaction hypothesis generation method can also be implemented by a processor calling computer-readable instructions stored in memory.

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0031] Please see the appendix Figure 1 The above is a flowchart illustrating a transaction hypothesis generation method based on semantic diffusion, as shown in an exemplary embodiment of this application. Figure 1 As shown, the transaction hypothesis generation method based on semantic diffusion in this embodiment may include the following steps S101~S104: S101: Obtain query task information for the target object; the target object is a transaction research target, and the query task information indicates query guidance information associated with the transaction research target.

[0032] The target object can be a trading research subject in the financial field. The types of trading research subjects can be various, such as a certain enterprise, company, stock, fund, or of course, a certain industry (such as the automotive industry or the catering industry) or a certain event (such as company A signing a cooperation contract with company B), etc. There are no restrictions here.

[0033] Here, the query task information indicates the query guidance information related to the trading research target. Specifically, it may include the target relationship type, query time window, holding period, risk constraints, and comparison benchmark.

[0034] Among them, the target association type is used to indicate the edge type that will be preferentially used in the subsequent diffusion process. For example, the target association type can be "supply chain relationship" or "industry association". The query time window is used to indicate the data time range considered by the semantic diffusion propagation query. The holding period refers to the expected holding period of the target object in the trading hypothesis generated based on the query task information, such as 5 trading days. The risk constraint refers to the risk control boundary that the trading hypothesis generated in the subsequent steps must adhere to in the backtesting process, such as profit-taking conditions, stop-loss conditions, position limits, etc. For example, the risk constraint is "maximum drawdown ≤ 8%". The benchmark is used to measure the reference benchmark for the performance of the trading hypothesis, such as the new energy vehicle exchange-traded open-ended index fund (or new energy ETF).

[0035] The query task information is shown in formula (1): (1) in, For the target object, For the target relationship type, For querying time windows, For the holding period, For risk constraints, This serves as the benchmark for comparison.

[0036] For example, the query task information for the target object could be: "Research which upstream and downstream companies will be affected if Company A experiences a supply chain disruption, look at the situation in the past six months, assume a holding period of 5 days, and keep the drawdown within 8%, benchmarking against a new energy ETF."

[0037] S102: Using the target object as a seed node, semantic diffusion propagation query is performed in the financial time heterogeneous graph with the query task information as semantic guidance to obtain multiple target nodes associated with the seed node; the financial time heterogeneous graph includes multiple types of nodes and multiple edges with multiple association types connecting different nodes; each node and each edge has a time label; the multiple types of nodes include multiple types of object nodes, information nodes, attribute nodes and decision nodes; the multiple association types include multiple types of temporal association, hierarchical association, causal association, information association and logical association.

[0038] Here, the financial time heterogeneous graph is a time-series heterogeneous graph structure, which includes multiple types of nodes and multiple edges with various types of relationships connecting different nodes. The time label refers to the time attribute information carried by each node and each edge, which is used to indicate the occurrence time of the node or the effective time of the relationship corresponding to the edge. It is the data foundation for realizing causal constraints and time-series filtering.

[0039] Among them, there are various types of nodes, including object nodes, information nodes, attribute nodes, and decision nodes. Object nodes are used to represent entities or transaction targets in the real world, such as asset nodes and company nodes; information nodes are used to represent information or event carriers, such as news nodes, announcement nodes, and event nodes; attribute nodes are used to represent meta-information describing the state of an entity or external conditions, such as industry nodes and macro nodes; and decision nodes are used to represent nodes that generate reasoning conclusions or historical decision records, such as hypothesis nodes and strategy nodes.

[0040] Among them, temporal correlation is used to indicate the sequential or temporal proximity relationship between different nodes in the time dimension, specifically including the sequence of events.

[0041] Hierarchical relationships are used to indicate structural hierarchical relationships between nodes, such as subordination, affiliation, upstream and downstream, or competition within the same industry. These include industry affiliation, supply chain relationship, and peer relationship.

[0042] Causal relationships are used to indicate the direct or indirect effects, transmission of influence, or risk mapping relationships that one node has on another node. Specifically, they include influence relationships, risk exposure relationships, and historical resonance relationships.

[0043] Information association is used to indicate the publishing, disclosure, or transmission relationship between information source nodes and subject nodes, specifically including publishing relationships.

[0044] Logical associations are used to indicate inference mappings, verification mappings, or reference standard associations between derived nodes and source nodes, specifically including hypothesis mappings, policy mappings, and benchmark associations.

[0045] For example, let's illustrate the 11 edge types one by one through a complete storyline (Company A → Industry → Upstream → Downstream → Macro → Decision): The story background is May 20, 2026, when Company A issues an announcement declaring "mass production of a new generation of sodium-ion batteries." This event triggers a series of cascading effects, which are recorded by the system as multiple edges.

[0046] I. Example of a publishing relationship: Company A (company node) published the "Announcement on Mass Production of New Generation Sodium-ion Batteries" (announcement node).

[0047] II. Example of industry relationship: Company A (company node) belongs to the power equipment industry (industry node).

[0048] III. Example of Impact Relationship: The "Announcement on Mass Production of Sodium-ion Batteries" (information node) affects Company A's stock price (asset node). The release of this announcement impacted Company A's stock price.

[0049] IV. Example of a supply chain relationship: The edge between Company A (company node) and a lithium mining company (company node) is a supply chain relationship type, indicating that Company A is a downstream customer of the lithium mining company, and lithium ore is an upstream raw material for batteries.

[0050] V. Example of industry-related relationship: The edge type between Battery Company A (Company Node) and Battery Company B (Company Node) is industry-related, indicating that they both belong to the battery industry and are direct competitors.

[0051] VI. Example of Risk Exposure Relationship: The edge type between Company A's stock price (asset node) and the lithium price fluctuation risk factor (risk factor node) is a risk exposure relationship, indicating that Company A's stock price is highly sensitive to lithium price fluctuations (lithium price increase → cost increase → profit decrease).

[0052] VII. Example of historical resonance relationship: The edge type between Company A's stock price (asset node) and the net asset value of the New Energy ETF (asset node) is a historical resonance relationship, indicating that Company A's stock price and the trend of the New Energy ETF were highly positively correlated in the past (correlation coefficient 0.85).

[0053] 8. Example of event sequence: The edge type between "Lithium ore price exceeds 500,000 yuan / ton" (event node) and "Sodium-ion battery mass production announcement" (event node) is an event sequence relationship, indicating that the lithium ore price increase event occurred before the sodium-ion battery announcement (April 2026 → May 2026).

[0054] IX. Example of Hypothetical Mapping Relationship: The edge type between hypothesis H-001 (hypothetical node) and the evidence path: announcement → lithium mine price increase → industry prosperity → Company A → new energy ETF (evidence path) is a hypothetical mapping relationship, indicating that the basis for hypothesis H-001 "Company A will outperform the new energy ETF" is this evidence path.

[0055] 10. Example of a strategy mapping relationship: The edge type between strategy S-001 (strategy node) and hypothesis H-001 (hypothesis node) is a strategy mapping relationship. This means that strategy S-001 (backtesting program) is compiled based on hypothesis H-001. After the backtest is completed, the performance data of this strategy is recorded in the strategy node and associated with the hypothesis node through this edge.

[0056] XI. Example of Benchmark Correlation: The edge type between the New Energy ETF (asset node) and the CSI 300 Index (benchmark node) is a benchmark correlation, indicating that the benchmark of the New Energy ETF is the CSI 300 Index, which is used to measure whether the ETF outperforms the market.

[0057] Furthermore, in the financial time heterogeneous graph, semantic diffusion propagation query is performed with the query task information as semantic guidance to obtain multiple target nodes associated with the seed node. Here, semantic diffusion propagation query refers to a graph query process that uses the query task information as semantic guidance, starts from the seed node on the time-series heterogeneous graph, and propagates outward layer by layer along the time dimension, dynamically selecting nodes based on semantic matching degree. The specific implementation of semantic diffusion propagation query is "wavefront propagation", which is manifested as diffusion outward layer by layer starting from the seed node. Each layer selects target nodes from the newly encountered candidate nodes in the current layer to enter the next layer, while retaining the nodes already selected in the previous layer, until the preset stopping condition is met, and the set of target nodes most relevant to the query task is selected.

[0058] The following section introduces the process of constructing a financial time heterogeneous graph.

[0059] Multiple types of financial data are acquired, multiple objects are extracted from the multiple types of financial data, and the relationships between the multiple objects are determined. Each object is identified as a node, and edges are constructed based on the relationships between the objects. Time labels and confidence labels are added to each node and each edge, and the financial time heterogeneous graph is constructed based on the multiple nodes and multiple edges.

[0060] The various types of financial data include transaction data, object status data, event log data, relationship description data, and classification identification data.

[0061] Transaction data is used to characterize the market conditions of the traded asset, such as price series, trading volume, and benchmark index information; object status data is used to characterize the financial status and allocation of the traded asset, such as financial statements and institutional holdings data; event record data is used to characterize events that have occurred or information that has been disclosed, such as news, announcements, and records of macroeconomic events; relationship description data is used to characterize the business or collaborative relationships between different objects, such as supply chain relationship data; classification and identification data is used to characterize the industry affiliation or theme classification of objects, such as industry classification and theme tags.

[0062] Here, the object types of multiple objects can include transaction entities, information carriers, event entities, category tags, and derived results. Transaction entities can include asset nodes and company nodes, information carriers can include news nodes and announcement nodes, event entities can refer to event nodes (such as interest rate hike events and merger and acquisition events), category tags can include industry nodes and macro nodes (such as CPI), and derived results can include hypothesis nodes and strategy nodes.

[0063] Furthermore, the raw financial data is transformed into the basic elements of a graph structure: nodes and edges. By identifying objects in the data and their relationships, the topological structure of a financial temporal heterogeneous graph is constructed. The relationships between objects include at least: subordinate or categorical relationships (e.g., an object belongs to a certain category), business collaboration relationships (e.g., upstream and downstream supply), impact transmission relationships (e.g., the effect of an event on an asset), and benchmarking relationships (e.g., the correspondence between an asset and a benchmark).

[0064] Finally, time labels, source labels, tradability labels, and confidence labels are added to each node and each edge, and a financial time heterogeneous graph is constructed based on multiple nodes and multiple edges.

[0065] Once the financial time heterogeneity graph is constructed, it becomes a time-series heterogeneous graph. , where the set of nodes It includes various types such as asset nodes, company nodes, news nodes, announcement nodes, event nodes, industry nodes, macro theme nodes, hypothesis nodes, strategy nodes, and benchmark nodes.

[0066] edge set It includes various types such as release relationship, industry relationship, impact relationship, supply chain relationship, peer relationship, risk exposure relationship, historical resonance relationship, event sequence relationship, hypothesis mapping relationship, strategy mapping relationship, and benchmark relationship.

[0067] Tag attributes: Each node and edge is accompanied by a time tag, source tag, confidence tag, and tradability tag.

[0068] S103: Based on the multiple target nodes and the edges between the nodes, construct an evidence subgraph, and perform message aggregation on the evidence subgraph to obtain the evidence vector of the evidence subgraph.

[0069] In this step, based on the multiple target nodes and their edges selected in S102, the corresponding local graph structure is extracted from the financial time heterogeneous graph to construct an evidence subgraph. The evidence subgraph preserves the connection relationships and temporal information between the target nodes, forming one or more evidence chains related to the query task.

[0070] The evidence subgraph then undergoes message aggregation processing: the feature information of each node and edge in the evidence subgraph is encoded and fused to form a unified evidence vector. This evidence vector incorporates the overall semantic information of the evidence subgraph.

[0071] S104: Based on the evidence subgraph and the evidence vector, perform hypothesis generation processing on the target object to obtain the transaction hypothesis corresponding to the target object.

[0072] In this step, the evidence vector is used as input, and the structured information of the evidence subgraph is combined to perform hypothesis generation processing, outputting a structured transaction hypothesis corresponding to the target object.

[0073] The trading assumptions include information on the decision-making entity (target asset, trading direction), information on the decision initiation (trigger event type, entry point, expected holding period), information on decision constraints (profit-taking conditions, stop-loss conditions, position limit, risk constraints), and information on decision references (benchmark, assumption confidence, evidence path set).

[0074] For example, the trading assumptions are as follows: Event type: major cooperation announcement, target stock: target company stock, direction: long, entry point: opening of the next trading day, holding period: 5 trading days, stop loss: -3%, take profit: +8%, benchmark: corresponding industry ETF, evidence path: announcement → partner → industry theme → historical similar events.

[0075] In this embodiment, semantic diffusion propagation query is performed on a temporal heterogeneous graph by using query task information as semantic guidance. The propagation process is dynamically filtered based on semantic matching degree, which significantly reduces the introduction of irrelevant nodes, improves the accuracy and precision of evidence discovery, and overcomes the static defects of fixed hop count search or keyword retrieval in the prior art.

[0076] Secondly, each node and side of the financial time heterogeneous graph has a time label, which allows the propagation process to be filtered and controlled according to the time dimension.

[0077] Finally, the above method realizes the process of query retrieval, semantic diffusion propagation, evidence subgraph extraction and aggregation, and structured hypothesis output. It can automatically complete the entire process transformation from raw query input to executable decision hypothesis without manual information organization or rule writing, thus bridging the technical gap between graph reasoning and structured output.

[0078] Please see Figure 2 This is a flowchart illustrating a semantic diffusion propagation query provided as an exemplary embodiment of this application. For example... Figure 2 As shown, step S102 includes steps S1021 to S1025: S1021: Using the target object as a seed node and the query task information as semantic guidance, perform multiple rounds of diffusion queries.

[0079] As can be seen from the preceding text, the semantic diffusion propagation query in this application is a wavefront propagation. Therefore, the query process is to start from the seed node and expand the query outward layer by layer on the financial time heterogeneous graph. Each round corresponds to a propagation layer (that is, the range of the seed node extending outward one step). Then, in each round, the newly added nodes newly contacted in the current layer are evaluated and screened.

[0080] During the query process, the query task information serves as the semantic guide. Specifically, the target relationship type in the query task information serves as the propagation direction guide. When expanding the query, edges that match the target relationship type are prioritized for propagation, while non-matching edge types are significantly suppressed. In this way, the propagation path can be dynamically adjusted according to different query intentions, rather than blindly traversing the graph.

[0081] S1022: For the current round of diffusion query, obtain the propagation node set obtained from the previous round of diffusion query, wherein the propagation node set includes at least one candidate node.

[0082] The current round refers to the round that is currently being executed. Round diffusion query, corresponding propagation depth (Extending outwards from the seed node) (Step), the previous round refers to the first... The scope of the diffusion query is as follows: This step provides a starting point for the current round of dissemination.

[0083] Propagation node set ( This refers to the set of nodes that were filtered and retained after the previous round of diffusion query. Each node in this set has undergone a comprehensive evaluation to confirm that it has sufficient relevance to the query task information. In wavefront propagation, the nodes selected in the previous layer are retained as the starting point (source node) for the next round of propagation.

[0084] Candidate nodes refer to nodes that are included in the evaluation scope in the current round of diffusion query but whose selection has not yet been determined. Based on these candidate nodes, the algorithm expands outward to find new neighboring nodes.

[0085] S1023: For each candidate node, determine the set of adjacent nodes of the candidate node, and remove the nodes that have been queried in previous rounds from each set of adjacent nodes to obtain the set of new nodes.

[0086] Adjacent node set This refers to the set of all nodes reached by extending outward one step along the edge from each candidate node in the previous round of propagation. The adjacent node set contains all new nodes that may participate in this round of propagation. If there are multiple candidate nodes in the previous round of propagation, then each candidate node corresponds to an adjacent node set.

[0087] Nodes that have been queried in each historical round That is, from the first It's the turn of the first The union of all nodes visited between rounds. The scope of history culling is controlled by the propagation range parameter. Constraints, only remove the most recent Nodes visited within a round, especially earlier ones, can re-enter the candidate set. This parameter is used in wavefront propagation to control the propagation window, ensuring that the flexibility to re-evaluate historical nodes is maintained while exploring new regions.

[0088] New Node Set This refers to the new set of nodes remaining after removing nodes that have been queried in previous rounds from each set of adjacent nodes. The new set of nodes is the candidate set that is eligible to be evaluated and filtered in the current round.

[0089] Please refer to formula (2), which is the expression for determining the propagation node in the current round: (2) in, For the newly added set of nodes, For the first The set of propagation nodes in each round, For the set of adjacent nodes, These are the nodes that have been queried in each historical round. The propagation window parameter is used to control the number of replay cycles (how many layers of historical nodes to replay). This is the set difference operator.

[0090] Here, by removing nodes that have been queried in previous rounds from the adjacent node set, it is ensured that the propagation wavefront always advances within the range of unexplored new nodes, avoiding repeated loops within the same node set. Propagation range control parameters. The propagation window was further constrained, achieving a balance between computational complexity and coverage in the propagation process. The smaller the size, the more focused the dissemination is on recently emerging areas. The larger the number of nodes, the more the propagation tends to global deduplication and extensive exploration. At the same time, the design of allowing earlier nodes to re-enter the candidate set retains the flexibility to re-evaluate historical nodes and avoids permanently losing potentially effective paths due to overly strict screening in a single instance.

[0091] Furthermore, each round of propagation node set includes nodes retained from the previous layer, and propagation always expands outward from the nodes retained in the previous round. This "retention + addition" set update mechanism ensures the continuity of the propagation path and the inter-layer dependencies: the propagation source nodes of each layer come from the screening results of the previous layer, making the final propagation path a coherent and traceable chain of evidence, rather than a scattered set of isolated nodes. The complete evidence path from the seed node to the target node is preserved, providing direct attribution traceability data for the evidence path set field in subsequent transaction hypotheses.

[0092] S1024: For each newly added node in the set of newly added nodes, determine the diffusion probability of the newly added node, and determine the propagation node of the current round from the set of newly added nodes in the current round based on the diffusion probability of each newly added node.

[0093] Diffusion probability ( (This refers to adding a new node) In the current round The probability of being selected and retained for the next round of propagation.

[0094] When determining the propagation nodes for the current round from the set of newly added nodes in the current round, the filtering method can be to select several nodes with the highest diffusion probability (Top-K sampling).

[0095] Specifically, regarding step S1024, when determining the diffusion probability of each newly added node in the set of newly added nodes, the following steps (i) to (ii) can be used: (i) For each newly added node, determine the diffusion coefficient of the newly added node; the diffusion coefficient is used to characterize the comprehensive correlation between the newly added node and the query task information.

[0096] diffusion coefficient It is in the During round propagation, under the current query task information, from the source node Spread to new nodes The overall relevance score. The larger the diffusion coefficient, the higher the relevance between the new node and the query task.

[0097] Specifically, when determining the diffusion coefficient of each newly added node, multiple diffusion index factors of the newly added node can be determined, and the diffusion coefficient can be determined based on the multiple diffusion index factors and their corresponding diffusion weight parameters.

[0098] The multiple diffusion index factors include a relevance factor, a quality factor, and an empirical factor. The relevance factor is used to measure the overall relevance between the new node and the query task; the quality factor is used to measure the quality level of the information carried by the new node, addressing the question of "whether it is reliable"; and the empirical factor is used to introduce feedback information from historical backtesting, addressing the questions of "whether it was effective and safe in the past". The relevance factors include semantic matching degree, relational matching degree, industry transmission coefficient, historical resonance strength, and risk exposure matching degree. The quality factors include time decay factor and information source credibility, while the empirical factors include historical backtesting contribution and future leakage penalty.

[0099] Among them, semantic matching degree refers to the degree of semantic relevance between the new node and the query task information; relation matching degree refers to whether the type of the edge connected to the new node matches the type of target association relationship in the query task information to ensure that the propagation proceeds in the correct direction; time decay factor refers to the timeliness between the time tag of the new node and the target query time window, with newer information receiving higher weight and older information having lower weight; information source credibility refers to the reliability level of the information source corresponding to the new node, with information from authoritative sources receiving higher weight; industry transmission strength is used to measure the transmission strength at the industry level, historical resonance strength is used to measure historical linkage, and risk exposure similarity is used to measure the consistency of risk dimensions; historical backtesting contribution refers to the performance contribution of the evidence path corresponding to the new node in historical backtesting, with historically effective paths receiving higher weight; future leakage penalty refers to whether the time tag of the new node is later than the deadline of the target query time window to prevent the use of future data in backtesting.

[0100] Furthermore, when determining the correlation factors, the following steps (1) to (7) may be included: (1) From the set of propagation nodes obtained in the previous round, determine the source node that is directly connected to the newly added node, and determine the relationship type of the edge between the source node and the newly added node.

[0101] From the previous round of propagation node set In the middle, find the newly added node. Directly connected nodes As the source node for the current propagation, and determine the connecting nodes. With nodes edge The relationship type.

[0102] (2) Determine the semantic matching degree based on the similarity between the newly added node and the query task information.

[0103] Input the associated text information of the newly added node (such as news title, announcement summary, node name, etc.) into the semantic coding model to obtain the first vector; input the query task information into the same semantic coding model to obtain the second vector; calculate the similarity (such as cosine similarity) between the first vector and the second vector as the semantic matching degree.

[0104] For example, if the query task information is "research on supply chain transmission risks" and the newly added node is "news on rising lithium prices", the two are highly related semantically, with a semantic matching degree of 0.92; if the newly added node is "news on company team building activities", the semantic matching degree is extremely low.

[0105] (3) Determine the relationship matching degree based on the relationship type of the edge between the newly added node and the source node and the matching degree between the target association relationship.

[0106] Relationship matching degree measures whether the path is correct, that is, whether the propagation is carried out in the direction specified by the user. For example, if the target relationship type in the query task information is "supply chain relationship", and the edge connecting the source node and the new node is also "supply chain relationship", then the relationship matching degree is 1.0; if it is "industry-related relationship", then the relationship matching degree is 0.1.

[0107] (4) Determine the industry transmission coefficient between the industry of the newly added node and the industry of the corresponding source node.

[0108] Here, the industry to which a newly added node belongs (e.g., the power equipment industry) can be read from the node attributes of the newly added node, and the industry to which a source node belongs (e.g., the non-ferrous metals industry) can be read from the node attributes of the source node. Based on the preset inter-industry transmission coefficient matrix, the transmission strength coefficient (i.e., the industry transmission coefficient) between the two industries can be determined.

[0109] (5) Determine the historical resonance intensity between the newly added node and the source node within a historical time period.

[0110] If the newly added node has a strong resonance relationship with the source node in the past (such as the stock price fluctuation of a supplier being highly correlated with the stock price fluctuation of a downstream customer), then when the source node is activated, the newly added node will have a higher diffusion probability, which will help capture the transmission path with historical linkage characteristics; if there is no significant resonance relationship in the past, the bonus will be smaller.

[0111] (6) Determine the degree of risk exposure matching between the newly added node and the risk constraint.

[0112] Specifically, the risk exposure labels of newly added nodes can be read from the financial time heterogeneity graph, and the degree of risk exposure matching can be determined based on the matching degree between risk constraints and risk exposure labels.

[0113] Risk exposure matching ensures that the selected nodes are compatible with the query task information in terms of risk dimension, avoiding the inclusion of nodes that conflict with risk control requirements in the propagation path.

[0114] (7) Determine the correlation factor based on the semantic matching degree, the relation matching degree, the industry transmission coefficient, the historical resonance intensity and the risk exposure matching degree.

[0115] Once the above indicators are determined, the correlation factors can be identified.

[0116] In some implementations, the determination of the quality factor can be achieved through the following steps (A) to (C): (A) Determine the time decay factor based on the time difference between the timestamp of the newly added node and the target query time window.

[0117] Here, the time difference between the newly added node's time label and the cutoff time of the target query time window is determined. Time difference Input exponential decay function The time decay factor is obtained.

[0118] in, The decay rate is controlled by a preset decay rate. The larger the time difference, the smaller the decay factor.

[0119] (B) Determine the credibility of the information source based on the confidence labels of the newly added nodes recorded in the financial time heterogeneity graph.

[0120] As mentioned earlier, nodes in a financial time heterogeneous graph are constructed with source and confidence level labels. The confidence level label is preset based on the authority level of the data source: the first data source (such as exchange announcements or institutional disclosures) has the highest confidence level (e.g., 0.95), the second data source (such as news media) has the next highest confidence level (e.g., 0.7), and the third data source (such as self-media) has the lowest confidence level (e.g., 0.4). In this way, the confidence level value can be directly read from the node attributes as the credibility of the information source.

[0121] For example, the newly added node is "Shanghai Stock Exchange Announcement Node" with a confidence level of 0.98; the newly added node is "A Financial Media Node" with a confidence level of 0.35.

[0122] (C) Determine the quality factor based on the time decay factor and the credibility of the information source.

[0123] After determining the time decay factor and the credibility of the information source, the quality factor can be determined based on the time decay factor and the credibility of the information source. Specifically, the quality factor can be obtained by weighted summation of the time decay factor and the credibility of the information source.

[0124] In some implementations, determining the empirical factors may include the following steps (a) to (c): (a) Determine the historical backtesting contribution based on the performance contribution of the evidence path corresponding to the newly added node in the historical backtesting.

[0125] Here, if the newly added node has appeared in a certain evidence path in the past, and that path has been backtested and verified to have made a positive performance contribution (such as excess returns, high Sharpe ratio, etc.), then the historical backtesting contribution score of that path is read from the historical backtesting records. If the node has no historical records, then a preset initial value or neutral value is used.

[0126] (b) Determine future leakage penalties based on the relationship between the timestamps of the newly added nodes and the backtesting deadline.

[0127] In this step, if the timestamp of a newly added node is later than the backtesting deadline, the node contains future information. Using it during backtesting will result in a future function. A positive penalty value (i.e., the penalty term is subtracted from the diffusion coefficient) is applied to such nodes. The magnitude of the penalty term is positively correlated with the extent to which the node's timestamp exceeds the backtesting deadline; the longer the delay, the greater the penalty. If the node's timestamp is not later than the backtesting deadline, the penalty term is zero.

[0128] In this way, future data leaks can be prevented from the source. Even if the candidate node is highly related to the query task, if it involves information generated after the backtesting point, the probability of its spread will be greatly reduced, ensuring that the backtesting results are true and reliable.

[0129] (c) Determine the empirical factor based on the historical backtesting contribution and the future leakage penalty.

[0130] After determining the historical backtesting contribution and the future leakage penalty, the empirical factor can be obtained. Specifically, the empirical factor can be determined based on the difference between the historical backtesting contribution and the future leakage penalty.

[0131] In this embodiment, please refer to formula (3), which is the function for determining the diffusion coefficient: (3) in, Where is the diffusion coefficient. For semantic matching degree, For relationship matching degree, The time decay factor, To assess the credibility of the information source, For industry transmission strength, For historical resonance intensity, For risk exposure similarity, Contribution to historical backtesting For future penalties for leaks, For diffusion weight parameters.

[0132] (ii) Based on the diffusion coefficient and the preset temperature parameter, determine the diffusion probability of the newly added node.

[0133] Diffusion probability It is in the During the round of propagation, query task information in the current stage. Below, add a new node The probability of being selected and retained for the next round of propagation is shown in formula (4): (4) in, Let be the diffusion probability. Where is the diffusion coefficient. This is the preset temperature parameter.

[0134] S1025: Add the propagation node of the current round to the propagation node set determined in the previous round until the preset stopping condition is met, and determine the node in the propagation node set as the target node.

[0135] Add the propagation nodes selected in the current round to the propagation node set determined in the previous round, forming an updated propagation node set. The propagation nodes in the current round. Then, it is determined whether a preset stopping condition is met. This could be reaching the maximum number of layers (maximum rounds), the new node set being empty (no new nodes for further propagation), the probability value of the node with the highest propagation probability being lower than a preset threshold, or the total number of nodes in the propagation node set reaching a preset threshold. If the stopping condition is met, propagation terminates, and all nodes in the current propagation node set are identified as target nodes. If not, the next round begins, and S1022 to S1025 are repeated.

[0136] Please see Figure 3 This is a flowchart illustrating another transaction hypothesis generation method based on semantic diffusion, provided as an exemplary embodiment of this application. Figure 1 Compared to the method shown, this method further includes steps S105-S107: S105: Compile the transaction hypothesis into a backtesting program.

[0137] This refers to using a pre-defined compilation language to convert trading assumptions into executable code.

[0138] In this invention, "compilation" specifically refers to mapping the 12 structured fields of the trading assumptions (such as trigger event type, target asset, entry point, profit-taking and stop-loss conditions, etc.) to specific logical modules in the backtesting program. Specifically, the backtesting program includes at least the following five built-in modules: event detector (listens to data stream and matches trigger conditions), order trigger (calculates slippage and order price), position controller (manages fund balance and position quantity), risk rule executor (independently monitors and enforces risk control rules such as stop-loss and position limit, with higher priority than strategy logic), and benchmark comparator (passively records the difference sequence between the strategy net value and the benchmark net value).

[0139] The backtesting program receives historical financial data as input and outputs the performance results of the strategy within that historical period.

[0140] S106: Obtain historical financial data within the target query time window and preset causal constraints; the causal constraints include at least one of time-series validity constraints, execution feasibility constraints, backtesting signal consistency constraints, and information completeness constraints.

[0141] It can be understood that historical financial data refers to financial data that has occurred within the query time window, including at least one of the following: historical market data (price series, trading volume), historical financial data, historical news data, historical announcement data, historical event data, and historical macroeconomic data.

[0142] The core characteristic of historical financial data is "time-series unidirectionality," meaning that the time label corresponding to the data must be earlier than or equal to the time point when the backtest is executed, to ensure that future data is not used during the backtest.

[0143] Timing validity constraints are used to ensure that trading signals are triggered only under conditions that comply with market time rules. Specifically, these constraints include: mapping between news release time and tradable time (after-hours announcements must be postponed to the opening of the next trading day), classification and processing of pre-market / intraday / after-hours news, and postponement on non-trading days (if the signal trigger date is a non-trading day, it will be postponed to the next trading day).

[0144] Feasibility constraints are used to ensure that trading signals are physically feasible to execute in a real market environment. These constraints include: position and capital constraints (verifying whether available funds in the account are sufficient to cover the order amount and transaction fees before placing an order; rejecting orders when funds are insufficient to prevent falsely high returns based on "unlimited ammunition"), transaction fee and slippage simulation (dynamically calculating impact costs based on order size and underlying asset liquidity; slippage for large orders is significantly greater than for small orders, preventing inflated returns due to the assumption of unlimited liquidity), and delisting and missing data handling (forced liquidation for trading suspensions exceeding the permitted period; returning null for missing data to prevent the use of future data for valuation).

[0145] Backtesting signal consistency constraints are used to ensure the consistency of execution logic when multiple trading signals are triggered simultaneously. Specifically, these constraints include: handling multiple signal conflicts (when multiple conflicting trading signals exist at the same time, the execution order is determined according to the preset priority of "risk control closing > take profit and stop loss > new opening position"), to avoid program deadlock or repeated opening and closing of positions to generate transaction fees.

[0146] Information completeness constraints are used to ensure that the backtesting process uses only publicly available and complete information at the decision-making time, preventing future data leakage from the source. Specifically, these constraints include: synchronous calculation of benchmark returns (synchronously calculating the concurrent returns of a preset benchmark for use in excess return calculations) and future information masking (the backtesting program is prohibited from using historical financial data generated after the backtesting time).

[0147] S107: Under the causal constraint, the historical financial data is backtested using the backtesting program to obtain the backtesting results.

[0148] Here, backtesting refers to the process of simulating the operation of a trading strategy on historical financial data to evaluate the strategy's performance over a past period.

[0149] The strategy performance evaluation report output after the backtesting program is completed. The backtesting results should include at least one of the following: annualized return (compound geometric return, not simple average), annualized volatility (annualized standard deviation of daily returns), Sharpe ratio (annualized strategy return minus risk-free rate divided by annualized volatility), maximum drawdown (maximum drop from peak to trough and corresponding duration), hit rate (the proportion of profitable trades to total trades after deducting fees), excess return (annualized result of daily subtraction between the strategy return curve and the benchmark return curve), and stability indicators (autocorrelation and win rate stability of the return series).

[0150] The above steps automatically compile the structured trading hypothesis into a backtesting program containing five built-in modules (event detector, order trigger, position controller, risk rule executor, and benchmark comparator). This solves the technical gap problem in existing technologies where graph reasoning results stop at natural language interpretation and cannot be directly used for backtesting verification. It enables evidence discovery results to automatically enter the verification process without the need for manual strategy code writing.

[0151] In other implementations, after obtaining backtesting results, the backtesting results are not only used to evaluate the merits of the strategy, but also serve as input data for feedback updates, so as to realize a feedback loop of evidence discovery, hypothesis generation, strategy compilation, causal backtesting, and parameter updates. Through continuous accumulation of backtesting experience, the strategy is continuously optimized, making the evidence paths generated in the future more and more accurate and more in line with the effectiveness of the real market.

[0152] Specifically, it can be determined whether the backtest result meets the preset requirements. If not, the diffusion weight parameter is adjusted based on the backtest result to obtain the adjusted diffusion weight parameter. Then, the step of determining the diffusion coefficient based on the multiple diffusion index factors and their corresponding diffusion weight parameters is returned to be executed until the backtest result meets the preset requirements.

[0153] Here, "determining whether the backtesting results meet the preset requirements" can mean comparing the "true value" of the backtesting results with the preset requirements to determine whether further iteration is needed. In practice, the preset requirements can be quantitative indicators such as return thresholds, Sharpe ratio thresholds, and maximum drawdown limits.

[0154] For example, when a certain type of evidence path consistently contributes positive excess returns in rolling window backtesting, the diffusion weight of that path is increased; when a certain type of path is often accompanied by time travel risk or overfitting, the penalty term is increased.

[0155] In this step, by using the backtesting results (actual performance calculated based on real historical financial data) as feedback signals, the system can automatically adjust the diffusion parameters according to the actual performance of the evidence path, so that effective paths are strengthened and ineffective paths are suppressed, realizing an end-to-end closed loop of "evidence discovery → strategy verification → propagation update", which is different from the one-time static mode of traditional systems where "backtesting is the end point".

[0156] It should be noted that, in addition to the aforementioned financial field, the method proposed in this application—"acquiring queries, navigating the graph, filtering nodes, aggregating information, and outputting structured results"—is essentially a general graph data processing and evidence discovery framework. This framework can also be applied to other fields, such as the medical field: constructing a temporal heterogeneous graph in the medical field, and then, for a symptom, querying related diseases and drugs, and outputting hypotheses about treatment plans. For another example, it can also be applied to the legal field: constructing a temporal heterogeneous graph in the legal field, and for a case, searching for relevant legal provisions and precedents, and outputting hypotheses about litigation strategies.

[0157] Corresponding to the aforementioned embodiments of the semantic diffusion-based transaction hypothesis generation method, this application also provides embodiments of a semantic diffusion-based transaction hypothesis generation apparatus.

[0158] Please see Figure 4 This is a schematic diagram illustrating the structure of a transaction hypothesis generation device based on semantic diffusion, as shown in an exemplary embodiment of this application. Figure 4 As shown, the semantic diffusion-based transaction hypothesis generation device 400 includes: The information acquisition module 410 is used to acquire query task information for a target object; the target object is a transaction research target, and the query task information indicates query guidance information associated with the transaction research target; The semantic query module 420 is used to take the target object as a seed node and perform semantic diffusion and propagation queries in the financial time heterogeneous graph with the query task information as semantic guidance to obtain multiple target nodes associated with the seed node; the financial time heterogeneous graph includes multiple types of nodes and multiple edges with multiple association types connecting different nodes; each node and each edge has a time label; the multiple types of nodes include multiple types of object nodes, information nodes, attribute nodes, and decision nodes; the multiple association types include multiple types of temporal association, hierarchical association, causal association, information association, and logical association. The message aggregation module 430 is used to construct an evidence subgraph based on the multiple target nodes and the edges between the nodes, and to perform message aggregation on the evidence subgraph to obtain the evidence vector of the evidence subgraph. Assume that model 440 is used to perform hypothesis generation processing on the target object based on the evidence subgraph and the evidence vector to obtain the transaction hypothesis corresponding to the target object.

[0159] In some implementations, the semantic query module 420 is specifically used for: Using the target object as a seed node and the query task information as semantic guidance, multiple rounds of diffusion queries are performed. For the current round of diffusion query, obtain the propagation node set obtained from the previous round of diffusion query, wherein the propagation node set includes at least one candidate node; For each candidate node, determine the set of adjacent nodes of the candidate node, and remove the nodes that have been queried in previous rounds from each set of adjacent nodes to obtain the set of new nodes; For each newly added node in the set of newly added nodes, determine the diffusion probability of the newly added node, and based on the diffusion probability of each newly added node, determine the propagation node of the current round from the set of newly added nodes in the current round. The propagation node of the current round is added to the propagation node set determined in the previous round until the preset stopping condition is met, and the node in the propagation node set is determined as the target node.

[0160] In some implementations, the semantic query module 420 is specifically used for: For each newly added node, a diffusion coefficient is determined; the diffusion coefficient is used to characterize the comprehensive correlation between the newly added node and the query task information. Based on the diffusion coefficient and the preset temperature parameter, the diffusion probability of the newly added node is determined.

[0161] In some implementations, the semantic query module 420 is specifically used for: For each new node, multiple diffusion index factors are determined for the new node; the multiple diffusion index factors include correlation factors, quality factors and empirical factors. The diffusion coefficient is determined based on the diffusion index factor and its corresponding diffusion weight parameter.

[0162] In some implementations, the query task information includes the target association type and risk constraints; the semantic query module 420 determines the relevance factors through the following steps: For the newly added node, from the propagation node set obtained in the previous round, determine the source node that is directly connected to the newly added node, and determine the relationship type of the edge between the source node and the newly added node; The semantic matching degree is determined based on the similarity between the newly added node and the query task information; The relationship matching degree is determined based on the degree of matching between the relationship type of the edge between the newly added node and the source node and the target association relationship; Determine the industry transmission coefficient between the industry of the newly added node and the industry of the corresponding source node; Determine the historical resonance intensity between the newly added node and the source node over a historical period; Determine the degree of risk exposure matching between the newly added node and the risk constraint; The correlation factor is determined based on the semantic matching degree, the relational matching degree, the industry transmission coefficient, the historical resonance intensity, and the risk exposure matching degree.

[0163] In some implementations, the query task information includes a target query time window, and the quality factor includes a time decay factor and information source credibility; the semantic query module 420 determines the quality factor through the following steps: The time decay factor is determined based on the time difference between the timestamp of the newly added node and the target query time window; The credibility of the information source is determined based on the confidence labels of the newly added nodes recorded in the financial time heterogeneity graph. The quality factor is determined based on the time decay factor and the credibility of the information source.

[0164] In some implementations, the semantic query module 420 determines the empirical factor through the following steps: The contribution of the new node to the evidence path in the historical backtesting is determined based on its performance contribution. Based on the relationship between the timestamps of the newly added nodes and the backtesting deadline, future leakage penalties will be determined. The empirical factor is determined based on the historical backtesting contribution and the future leakage penalty.

[0165] Please see Figure 5 This is a schematic diagram illustrating the structure of another transaction hypothesis generation device based on semantic diffusion, as shown in an exemplary embodiment of this application. Figure 5 As shown, the semantic diffusion-based transaction hypothesis generation device 400 also includes a backtesting module 450; the query task information includes a target query time window; the backtesting module 450 is used for: Compile the aforementioned transaction assumptions into a backtesting program; Obtain historical financial data within the target query time window and preset causal constraints; the causal constraints include at least one of the following: time series validity constraints, execution feasibility constraints, backtesting signal consistency constraints, and information completeness constraints. Under the aforementioned causal constraints, the historical financial data is backtested using the backtesting procedure to obtain backtesting results.

[0166] In some embodiments, the backtesting module 450 is further configured to: Determine whether the backtest results meet the preset requirements. If not, adjust the diffusion weight parameters based on the backtest results to obtain the adjusted diffusion weight parameters. Return to the step of determining the diffusion coefficient based on the multiple diffusion index factors and their corresponding diffusion weight parameters, until the backtesting results meet the preset requirements.

[0167] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0168] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0169] Corresponding to the above-described semantic diffusion-based transaction hypothesis generation method, this disclosure also provides a computer device, such as... Figure 6The diagram shown is a structural schematic of a computer device provided in an embodiment of this disclosure. Figure 6 As shown, the computer device 600 includes a processor 610, an internal bus 620, memory 630, a network interface 640, and non-volatile memory 650, and may also include other hardware required for its functions. One or more embodiments of this specification can be implemented in software, for example, the processor 610 reads the corresponding computer program from the non-volatile memory 650 into the memory 630 and then runs it. Of course, besides software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0170] The memory 630, also known as internal memory, is used to temporarily store the computational data in the processor 610, as well as the data exchanged with non-volatile memory 650 such as hard disk. The processor 610 exchanges data with non-volatile memory 650 through the memory 630.

[0171] In this embodiment, memory 630 is specifically used to store application code that executes the solution of this application, and its execution is controlled by processor 610. That is, when the computer device is running, processor 610 communicates with network interface 640, memory 630 and non-volatile memory 650 through internal bus 620, so that processor 610 executes the application code stored in memory 630 and non-volatile memory 650, thereby executing the transaction hypothesis generation method based on semantic diffusion described in the above method embodiment.

[0172] Processor 610 may be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware microservices. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0173] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the computer device 600. In other embodiments of this application, the computer device 600 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0174] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the semantic diffusion-based transaction hypothesis generation method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0175] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the semantic diffusion-based transaction hypothesis generation method in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0176] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0177] The embodiments of the subject matter and functional operation described in this specification can be implemented in the following ways: digital electronic circuits, tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing apparatus or for controlling the operation of a data processing apparatus. Alternatively or additionally, the program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or combinations thereof.

[0178] The processing and logic flow described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flow can also be executed by dedicated logic circuitry—such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits), and the device can also be implemented as dedicated logic circuitry.

[0179] Computers suitable for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. Basic computer microservices include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0180] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. Processors and memory may be supplemented by or incorporated into dedicated logic circuitry.

[0181] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0182] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and microservices in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program microservices and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0183] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0184] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for generating transaction hypotheses based on semantic diffusion, characterized in that, include: Retrieve query task information for the target object; The target object is the transaction research target, and the query task information indicates query guidance information associated with the transaction research target; Using the target object as a seed node, semantic diffusion propagation query is performed in the financial time heterogeneous graph with the query task information as semantic guidance to obtain multiple target nodes associated with the seed node; the financial time heterogeneous graph includes multiple types of nodes and multiple edges with multiple association types connecting different nodes; each node and each edge has a time label; the multiple types of nodes include multiple types of object nodes, information nodes, attribute nodes and decision nodes. The various types of association relationships include multiple types such as temporal association relationships, hierarchical association relationships, causal association relationships, information association relationships, and logical association relationships; Based on the multiple target nodes and the edges between the nodes, an evidence subgraph is constructed, and message aggregation is performed on the evidence subgraph to obtain the evidence vector of the evidence subgraph; Based on the evidence subgraph and the evidence vector, hypothesis generation processing is performed on the target object to obtain the transaction hypothesis corresponding to the target object.

2. The method according to claim 1, characterized in that, The step involves using the target object as a seed node and performing semantic diffusion propagation queries in the financial time heterogeneous graph, guided by the query task information, to obtain multiple target nodes associated with the seed node, including: Using the target object as a seed node and the query task information as semantic guidance, multiple rounds of diffusion queries are performed. For the current round of diffusion query, obtain the propagation node set obtained from the previous round of diffusion query, wherein the propagation node set includes at least one candidate node; For each candidate node, determine the set of adjacent nodes of the candidate node, and remove the nodes that have been queried in previous rounds from each set of adjacent nodes to obtain the set of new nodes; For each newly added node in the set of newly added nodes, determine the diffusion probability of the newly added node, and based on the diffusion probability of each newly added node, determine the propagation node of the current round from the set of newly added nodes in the current round. The propagation node of the current round is added to the propagation node set determined in the previous round until the preset stopping condition is met, and the node in the propagation node set is determined as the target node.

3. The method according to claim 2, characterized in that, Determining the diffusion probability of each new node in the set of new nodes includes: For each newly added node, a diffusion coefficient is determined; the diffusion coefficient is used to characterize the comprehensive correlation between the newly added node and the query task information. Based on the diffusion coefficient and the preset temperature parameter, the diffusion probability of the newly added node is determined.

4. The method according to claim 3, characterized in that, Determining the diffusion coefficient of each newly added node includes: For each new node, multiple diffusion index factors are determined for the new node; the multiple diffusion index factors include correlation factors, quality factors and empirical factors. The diffusion coefficient is determined based on the multiple diffusion index factors and their corresponding diffusion weight parameters.

5. The method according to claim 4, characterized in that, The query task information includes the target relationship type and risk constraints; the relevance factor is determined through the following steps: For the newly added node, from the propagation node set obtained in the previous round, determine the source node that is directly connected to the newly added node, and determine the relationship type of the edge between the source node and the newly added node; The semantic matching degree is determined based on the similarity between the newly added node and the query task information; The relationship matching degree is determined based on the degree of matching between the relationship type of the edge between the newly added node and the source node and the target association relationship; Determine the industry transmission coefficient between the industry of the newly added node and the industry of the corresponding source node; Determine the historical resonance intensity between the newly added node and the source node over a historical period; Determine the degree of risk exposure matching between the newly added node and the risk constraint; The correlation factor is determined based on the semantic matching degree, the relational matching degree, the industry transmission coefficient, the historical resonance intensity, and the risk exposure matching degree.

6. The method according to claim 4, characterized in that, The query task information includes the target query time window, and the quality factor includes a time decay factor and the credibility of the information source; the quality factor is determined through the following steps: The time decay factor is determined based on the time difference between the timestamp of the newly added node and the target query time window; The credibility of the information source is determined based on the confidence labels of the newly added nodes recorded in the financial time heterogeneity graph. The quality factor is determined based on the time decay factor and the credibility of the information source.

7. The method according to claim 4, characterized in that, The empirical factor is determined through the following steps: The contribution of the new node to the evidence path in the historical backtesting is determined based on its performance contribution. Based on the relationship between the timestamps of the newly added nodes and the backtesting deadline, future leakage penalties will be determined. The empirical factor is determined based on the historical backtesting contribution and the future leakage penalty.

8. The method according to claim 4, characterized in that, The query task information includes a target query time window; after obtaining the transaction hypothesis corresponding to the target object, the method further includes: Compile the aforementioned transaction assumptions into a backtesting program; Obtain historical financial data within the target query time window and preset causal constraints; the causal constraints include at least one of the following: time series validity constraints, execution feasibility constraints, backtesting signal consistency constraints, and information completeness constraints. Under the aforementioned causal constraints, the historical financial data is backtested using the backtesting procedure to obtain backtesting results.

9. The method according to claim 8, characterized in that, After obtaining the backtest results, the method further includes: Determine whether the backtest results meet the preset requirements. If not, adjust the diffusion weight parameters based on the backtest results to obtain the adjusted diffusion weight parameters. Return to the step of determining the diffusion coefficient based on the multiple diffusion index factors and their corresponding diffusion weight parameters, until the backtesting results meet the preset requirements.

10. The method according to claim 1, characterized in that, The financial time heterogeneity graph is constructed using the following steps: Acquire various types of financial data; these include transaction data, object status data, event log data, relationship description data, and classification identification data. Multiple objects are extracted based on the various types of financial data, and the relationships between the multiple objects are determined; Each object is identified as a node, and edges are constructed based on the relationships between objects; Add a time label and a confidence label to each node and each edge, and construct the financial time heterogeneous graph based on the multiple nodes and multiple edges.

11. A transaction hypothesis generation device based on semantic diffusion, characterized in that, include: The information acquisition module is used to acquire query task information for the target object; The target object is the transaction research target, and the query task information indicates query guidance information associated with the transaction research target; The semantic query module is used to take the target object as a seed node and perform semantic diffusion and propagation queries in the financial time heterogeneous graph with the query task information as semantic guidance to obtain multiple target nodes associated with the seed node; the financial time heterogeneous graph includes multiple types of nodes and multiple edges with multiple association types connecting different nodes; each node and each edge has a time label; the multiple types of nodes include multiple types of object nodes, information nodes, attribute nodes and decision nodes; The various types of association relationships include multiple types such as temporal association relationships, hierarchical association relationships, causal association relationships, information association relationships, and logical association relationships; The message aggregation module is used to construct an evidence subgraph based on the multiple target nodes and the edges between the nodes, and to perform message aggregation on the evidence subgraph to obtain the evidence vector of the evidence subgraph. The hypothesis generation model is used to perform hypothesis generation processing on the target object based on the evidence subgraph and the evidence vector to obtain the transaction hypothesis corresponding to the target object.

12. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the transaction hypothesis generation method based on semantic diffusion as described in any one of claims 1-10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the semantic diffusion-based transaction hypothesis generation method according to any one of claims 1-10.

Images