Transaction counterparty recommendation method and system based on multi-modal data

Abstract

The application discloses a counterparty recommendation method and system based on multi-modal data, and relates to the technical field of financial software development. The method comprises the following steps: acquiring multi-modal data; performing data fusion on the multi-modal data, configuring multi-dimensional portrait labels for each candidate counterparty according to the fused data, including accurate labels and fuzzy labels; constructing a counterparty matrix including label information of all candidate counterparties according to the multi-dimensional portrait labels; when receiving an investment instruction, analyzing and obtaining transaction characteristic elements of a bond product to be traded; when collecting a quotation demand initiated by a trader for the bond product, determining a target counterparty matched with the bond product through a multi-dimensional matching model according to the aforementioned counterparty matrix, and then outputting target counterparty information and / or sending a quotation instruction to the target counterparty. The application can significantly improve the efficiency and success rate of bond quotation transactions.

Description

Technical Field

[0001] This invention relates to the field of financial software development technology, and in particular to a method and system for recommending trading counterparties based on multimodal data. Background Technology

[0002] With the increasing networking and globalization of the financial industry, and in response to rapid market changes and the business development of banks, transaction management systems for managing interbank market transaction data have emerged, such as the ComStar system. The ComStar system covers all business types of the China Foreign Exchange Trading Center's (CFETS) RMB and foreign currency trading platform, spot and repurchase transactions of the Shanghai and Shenzhen Stock Exchanges, spot and derivatives trading of the Shanghai Gold Exchange, discounting and repurchase transactions of the Shanghai Commercial Paper Exchange, treasury bond futures business of the China Financial Futures Exchange, and various offline businesses. It achieves integrated RMB and foreign currency processing and direct front-end, middle-end, and back-end processing, significantly reducing manual operations and lowering operational risks. Simultaneously, the ComStar system can seamlessly integrate with the CFETS trading platform and post-trade platform, enabling direct processing of trading strategies, pre-approval, real-time quota control and transaction confirmation, and fund clearing.

[0003] The interbank market and the exchange market are the two main venues for trading financial instruments such as bonds and foreign exchange. Currently, in certain bond transactions (such as credit bond transactions) in the interbank and exchange markets, the process is mainly conducted through bilateral negotiations and price inquiries. Traders use chat tools to inquire about prices from counterparties, and the choice of counterparty determines the efficiency and cost of completing the transaction. This process is highly dependent on the trader's personal experience and network of contacts, and has the following drawbacks:

[0004] 1) Inefficient: Traders need to manually screen and memorize a large number of counterparty preferences, which is time-consuming and labor-intensive.

[0005] 2) High subjectivity: The matching process relies on personal experience, which can easily lead to overlooking high-quality opponents or choosing non-optimal opponents.

[0006] 3) Inability to share resources: The resources held by each trader cannot be shared. New traders lack experience and find it difficult to start working quickly and effectively. When employees leave, they can easily take resources with them.

[0007] To address the aforementioned shortcomings, existing technologies offer several electronic trading platforms capable of intelligently recommending trading counterparties. A common approach is to use historical transaction data, employing traditional data analysis tools or simple rule engines to perform statistical analysis and then recommend relevant institutions (counterparties) to traders. However, the institution lists provided by these platforms typically only list relevant institutions (counterparties) and lack intelligent sorting and recommendation functions, leaving room for improvement in the intelligence and accuracy of counterparty recommendations. Furthermore, these solutions rely on static analysis of historical data, lacking dynamic data processing capabilities and failing to flexibly respond to dynamic changes in the trading market, potentially causing recommendations to lag behind actual market conditions.

[0008] On the other hand, with the development and popularization of information processing methods such as artificial intelligence, big data analytics, and machine learning, existing technologies offer various technical solutions for automated, personalized, and precise marketing decision-making and execution based on user behavior data, transaction data, and external environmental variables. Their core components typically include user profiling, behavioral prediction modeling, marketing strategy generation, marketing content delivery, and channel optimization. They emphasize using technological means to refine user identification and classification for differentiated marketing, primarily involving data collection and processing, model building and training, strategy decision generation, information delivery execution, and effect tracking and feedback analysis. This is a crucial component closely integrated with marketing promotion in digital enterprise operations. Based on these technologies, existing technologies also offer some intelligent marketing decision-making solutions applicable to the financial sector. For example, Chinese patent ZL202510954044.6 discloses a marketing method for financial service products based on DeepSeek, including the following steps: collecting user browsing time sequences of financial products and fund subscription interval sequences; calculating the volatility index generated by changes in attention; comparing the deviation between the predicted and real-time curves by combining policy hot words and yield data; optimizing the recommendation ranking based on the deviation sign and volatility range using reinforcement learning; matching personalized combinations of user and product feature vectors; monitoring the update time sequence of operational behavior and analyzing behavior scores. In the above scheme, the volatility index of financial behavior conversion is tracked through a dynamic window; the market deviation of hot words and yields is combined; reinforcement learning adjusts the recommendation ranking based on the deviation state; feature vector matching achieves accurate recommendations; and closed-loop feedback of behavior monitoring continuously optimizes the strategy, constructing an intelligent decision-making chain from data collection to strategy adjustment, breaking through the limitations of traditional analysis and recommendation, and improving the timeliness, accuracy, and adaptability of marketing. However, the above financial service product marketing scheme is difficult to directly apply to the inquiry and quotation trading of financial products.

[0009] Based on existing intelligent marketing decision-making solutions, and combined with the needs and characteristics of inquiry and trading of financial products (such as credit bonds), this invention provides a counterparty recommendation scheme that can significantly improve inquiry efficiency and success rate. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for recommending trading counterparties based on multimodal data. This invention first constructs dynamic, multi-dimensional profile tags for trading counterparties based on multimodal data. Upon receiving an investment instruction, the trader can initiate a price inquiry based on the bond information in the instruction. After the inquiry is initiated, multi-dimensional matching calculations can quickly identify target trading counterparties that match the bond products, and automatically send price inquiries to those target counterparties. This helps traders achieve accurate and rapid transactions. This invention can significantly improve the efficiency and success rate of bond price inquiry transactions.

[0011] To achieve the above objectives, the present invention provides the following technical solution:

[0012] A counterparty recommendation method based on multimodal data includes the following steps:

[0013] Acquire multimodal data, which is multi-source data related to bond trading obtained from different channels;

[0014] The acquired multimodal data is fused, and multi-dimensional profile labels are configured for each candidate counterparty based on the fused data. The multi-dimensional profile labels include precise labels and fuzzy labels. The precise labels are used to identify the bond codes that the candidate counterparty expects to trade, and the fuzzy labels are used to identify the bond preference characteristics that the candidate counterparty expects to trade.

[0015] Based on the multi-dimensional profile tags of each candidate counterparty, a counterparty matrix including tag information of all candidate counterparties is constructed;

[0016] Upon receiving an investment instruction, the system parses the instruction to obtain the transaction feature elements of the bond product to be traded, which include at least the bond code.

[0017] When a trader initiates a price inquiry for the aforementioned bond product, the system determines the target counterparty that matches the bond product based on the aforementioned counterparty matrix and a multi-dimensional matching model, and then outputs the target counterparty information and / or sends a price inquiry instruction to the target counterparty.

[0018] Furthermore, the candidate trading counterparties are obtained from the trader's contacts in the associated chat tool's address book. In this case, the constructed trading counterparty matrix is ​​a friend matrix.

[0019] Furthermore, the fuzzy label includes a preference label and a behavioral feature label. The preference label is used to identify the type of bonds preferred by the candidate counterparty during the transaction, and the behavioral feature label is used to identify the major categories of bonds preferred by the candidate counterparty during the transaction.

[0020] The bond type is distinguished by the sub-type of the bond issuer, and the bond category is distinguished by the parent type of the bond issuer, with one or more sub-types under the parent type.

[0021] Furthermore, the multimodal data includes structured data and unstructured data;

[0022] The structured data includes historical transaction data and bond data using a preset standard data format;

[0023] The unstructured data includes unstructured interactive data and unstructured real-time market data. The unstructured interactive data includes chat data from associated chat tools, and the unstructured real-time market data includes intended price quotes from associated chat tools.

[0024] Furthermore, the steps of configuring multi-dimensional profile tags for each candidate counterparty based on the fused data include:

[0025] For each candidate counterparty, real-time chat data and quote data from the associated chat tools are integrated and semantic analysis is performed. Based on the semantic analysis results, it is determined whether the candidate counterparty has the bond code for the desired transaction. If so, the bond code is used as the precise label for the candidate counterparty, and the date of the data occurrence is recorded.

[0026] Furthermore, after integrating the historical transaction data and bond data of the candidate counterparty, data semantic analysis is performed. Based on the semantic analysis results, it is determined whether the candidate counterparty has a preferred bond type. If so, the bond type is used as a preference label for the candidate counterparty, and the date of the data occurrence is recorded.

[0027] Furthermore, after integrating the historical transaction data and quotation data of the candidate counterparty, data semantic analysis is performed. Based on the semantic analysis results, it is determined whether the candidate counterparty has a preferred bond category. If so, the bond category is used as a behavioral characteristic label for the candidate counterparty, and the data occurrence date is recorded.

[0028] Furthermore, after configuring the multi-dimensional profile tags for each candidate counterparty, the aforementioned counterparty matrix is ​​constructed based on the candidate counterparty's name and / or number, multi-dimensional profile tags, data occurrence date, and data type.

[0029] The elements of the counterparty matrix include at least the name and / or number of the candidate counterparty, precise tags, preference tags, behavioral characteristic tags, data occurrence date, and data type information.

[0030] Furthermore, the matching degree between the candidate counterparty and the bond product is obtained by calculating the matching degree score. The multi-dimensional matching model is configured to use a hierarchical matching method that combines precise matching and association rule matching to calculate the matching degree score between the bond product and the candidate counterparty.

[0031] At this point, the bond code of the bond product is first matched with the precise tag to obtain the matching counterparty. When the number of precisely matched counterparties does not meet the preset number requirement, the tag association rule matching is triggered to obtain more counterparties that match the aforementioned bond product.

[0032] Furthermore, the steps for calculating the matching score between bond products and candidate counterparties using a tiered matching method are as follows:

[0033] S110, for each bond product in the investment instruction, the counterparty matrix is ​​queried based on the bond code of the bond product to obtain the precise label matching the bond code, and the candidate counterparty information to which these precise labels belong is extracted to form the counterparty information to be matched; for each counterparty to be matched, the weight of the precise label data of the counterparty to be matched is adjusted according to the time-related label factor parameter, and the matching degree score between the counterparty to be matched and the bond product is calculated by weighted algorithm according to the adjusted weight value; the matching degree scores of all counterparties to be matched are sorted from largest to smallest, thus completing the precise matching.

[0034] S120: Determine whether the number of counterparties to be matched has reached a preset threshold N, where N is an integer greater than or equal to 2. If the preset threshold N is reached, proceed to step S130. If the preset threshold N is not reached, proceed to step S140.

[0035] S130: Output the top N sorted counterparties to be matched as target counterparty recommendations, and then end.

[0036] S140, trigger association rule matching. At this time, after constructing a counterparty-bond dataset based on the precise and fuzzy label data in the counterparty matrix, the associated bond products with the highest correlation to the aforementioned bond products and appearing in the precise label data are obtained through a preset confidence algorithm. The aforementioned precise matching is performed through the bond codes of the associated bond products to obtain the target counterparties related to the associated bond products. The target counterparties of the associated bond products are then recommended and output as the target counterparties of the aforementioned bond products.

[0037] Furthermore, the time-related labeling factors include a time enhancement factor and a time decay factor; the weight value corresponding to the time enhancement factor increases over time within a preset time range; the weight value corresponding to the time decay factor decreases over time within a preset time range.

[0038] Furthermore, based on the data type information in the counterparty matrix, each data type is configured as an item, and a maximum score (max_score) is assigned to each data type item. The steps for calculating the matching score between the counterparty to be matched and the bond product using a weighted algorithm are as follows:

[0039] For each data type item, calculate the individual label score (label_score) for each precise label as follows: After determining the weight value corresponding to the label factor of the precise label based on the data occurrence date, multiply the weight value by the preset initial label score of the precise label to obtain the individual label score (label_score) of the precise label, i.e., label_score = label factor weight value * initial label score;

[0040] The tag scores of multiple precise tags belonging to the same data type item are summed to obtain the score item_score for each item. The value of item_score does not exceed the maximum score max_score of the data type item configuration item, that is, item_score = Min( (max_score);

[0041] The scores of multiple data type items related to the bond product are summed to obtain the total bond score, bond_score.

[0042] The value of bond_score does not exceed the pre-configured score threshold MAX, that is, bond_score = Min( ,MAX).

[0043] Furthermore, the confidence algorithm is the FP-Growth algorithm. In this case, the steps for obtaining the associated bond products with the highest correlation to the aforementioned bond products and appearing in the precise labeling data are as follows:

[0044] A counterparty-bond dataset is constructed based on the precise and fuzzy label data in the counterparty matrix. The counterparty-bond dataset records candidate counterparties and historical bond data purchased by the candidate counterparties.

[0045] Based on the frequency of bonds appearing in the counterparty-bond dataset, calculate the support for each bond; based on the preset minimum support, filter out bonds with support values ​​less than the minimum support.

[0046] The remaining bonds after filtering are sorted by support to obtain the sorted dataset;

[0047] Construct a frequent item tree (FP tree) based on the sorted dataset mentioned above.

[0048] Based on the constructed FP treee, the conditional pattern base data of the bond products in the investment instructions are obtained;

[0049] Based on the conditional model base data of bond products, and combined with the support tables of each bond, the confidence level of the relevant bonds is calculated.

[0050] Bonds are sorted by their confidence level, and the bonds with the highest confidence level that appear in the precise tag data are identified as related bond products. These related bond products are then precisely matched to obtain target counterparty information.

[0051] The present invention also provides a counterparty recommendation system for an electronic trading platform, the system comprising:

[0052] The data acquisition module is used to acquire multimodal data, which is multi-source data related to bond trading acquired from different channels;

[0053] The tag configuration module is used to perform data fusion on the acquired multimodal data and configure multi-dimensional profile tags for each candidate counterparty based on the fused data. The multi-dimensional profile tags include precise tags and fuzzy tags. The precise tags are used to identify the bond codes that the candidate counterparty hopes to trade, and the fuzzy tags are used to identify the bond preference characteristics that the candidate counterparty hopes to trade.

[0054] The matrix building module is used to construct a counterparty matrix that includes the tag information of all candidate counterparties based on the multi-dimensional profile tags of each candidate counterparty;

[0055] The investment instruction parsing module is used to parse the investment instruction upon receiving it and obtain the transaction feature elements of the bond product to be traded. The transaction feature elements include at least the bond code.

[0056] The counterparty matching module is used to collect price inquiry requests from traders for the aforementioned bond products, determine the target counterparty matching the bond products based on the aforementioned counterparty matrix and a multi-dimensional matching model, output the target counterparty information, and / or send price inquiry instructions to the target counterparty.

[0057] Compared with the prior art, this invention, by adopting the above technical solution, has the following advantages and positive effects: First, this invention constructs dynamic multi-dimensional profile tags for counterparties based on multimodal data. After receiving an investment instruction, the trader can initiate an inquiry based on the bond information in the investment instruction. After the inquiry is initiated, the target counterparty information matching the bond product can be quickly locked through multi-dimensional matching calculation, and an inquiry can be automatically sent to the target counterparty. In this way, it can help the trader to achieve a transaction accurately and quickly.

[0058] Specifically, on the one hand, this invention constructs a data matrix based on multimodal data, effectively solving the current problem of failing to recommend trading counterparties (friends) for inactive bonds. On the other hand, through a hierarchical matching algorithm, it reduces model illusions and improves the efficiency and accuracy of the algorithm. Furthermore, by introducing time enhancement and decay factors into the multi-dimensional matching calculation, it addresses the issue of changing trading intentions over time, thereby improving the accuracy of recommending friends.

[0059] This invention offers the following advantages: First, it improves transaction efficiency: transforming manual screening into second-level automatic recommendations, significantly reducing manpower requirements. Second, it increases transaction success rate: precise recommendations based on a dynamic data matrix significantly improve the hit rate of price inquiries. Third, it lowers the trading threshold: helping new traders get started quickly and narrowing the experience gap. Fourth, it provides adaptive optimization: it can adjust based on transaction results, continuously optimizing the recommendation method, becoming increasingly intelligent with use. Fifth, it reduces costs and increases efficiency: quickly and accurately recommending a large number of potential clients and automatically sending price inquiry instructions helps traders select the optimal price for execution. Attached Figure Description

[0060] Figure 1 The logical structure diagram of the counterparty recommendation method based on multimodal data provided in the embodiments of the present invention is shown.

[0061] Figure 2 This is an example diagram of the data structure of the friend matrix provided in an embodiment of the present invention.

[0062] Figure 3 This is an example diagram of label factor configuration provided in an embodiment of the present invention.

[0063] Figure 4 This is an example diagram of a friend-bond dataset provided in an embodiment of the present invention.

[0064] Figure 5 The filtered bond sorting chart provided for an embodiment of the present invention.

[0065] Figure 6 An example diagram of the sorted dataset provided in an embodiment of the present invention.

[0066] Figure 7The logical structure diagram of the FP tree provided in the embodiment of the present invention.

[0067] Figure 8 This is a schematic diagram of the conditional pattern base of bond G provided in an embodiment of the present invention. Detailed Implementation

[0068] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed account of the counterparty recommendation method and system based on multimodal data disclosed in this invention. It should be noted that techniques (including methods and apparatuses) known to those skilled in the art may not be discussed in detail, but where appropriate, such known techniques are considered part of the specification. Furthermore, other examples of exemplary embodiments may have different values. The structures, proportions, sizes, etc., depicted in the accompanying drawings are merely illustrative of the content disclosed in this specification for the understanding and reading of those skilled in the art, and are not intended to limit the conditions under which the invention can be implemented.

[0069] In the description of the embodiments of this application, " / " means "or", and "and / or" is used to describe the relationship between related objects, indicating that there can be three relationships. For example, "A and / or B" means: A and B exist alone, B exists alone, and A and B exist simultaneously. In the description of the embodiments of this application, "multiple" refers to two or more.

[0070] The technical concept and solution of the present invention will be described below based on exemplary application scenarios. Example

[0071] The core of this invention lies in constructing a dynamic, multi-dimensional counterparty profile system and a multi-dimensional matching system. When a trader initiates an inquiry, the system can quickly identify suitable counterparties for the bonds to be bought and sold, recommend these counterparties to the trader, and automatically send inquiry instructions for relevant bonds to these counterparties to help the trader achieve accurate and rapid transactions.

[0072] Specifically, this invention provides a counterparty recommendation method based on multimodal data, the method comprising the following steps:

[0073] S100, acquire multimodal data, which is multi-source data related to bond trading obtained from different channels. Specifically, the multimodal data refers to multi-channel data related to bond trading that differs in generation mechanism, data format, and information content.

[0074] S200: Data fusion is performed on the acquired multimodal data, and multi-dimensional profile labels are configured for each candidate counterparty based on the fused data. The multi-dimensional profile labels include precise labels and fuzzy labels. The precise labels identify the bond codes that the candidate counterparty intends to trade, and the fuzzy labels identify the bond preference characteristics that the candidate counterparty intends to trade. Preferably, the fuzzy labels include preference labels and behavioral feature labels. The preference labels identify the bond types preferred by the candidate counterparty during trading. The behavioral feature labels identify the major bond categories preferred by the candidate counterparty during trading. The bond types are distinguished by sub-types of the bond issuer, and the major bond categories are distinguished by parent types of the bond issuer, with each parent type having one or more sub-types.

[0075] S300 constructs a counterparty matrix that includes tag information for all candidate counterparties based on the multi-dimensional profile tags of each candidate counterparty.

[0076] S400, upon receiving an investment instruction, parses the instruction to obtain the trading characteristic elements of the bond product to be traded. These trading characteristic elements include at least the bond code. Specifically, the trading characteristic elements may include the bond code and basic characteristics, including but not limited to bond type (e.g., medium-term notes), issuer (e.g., **Company), industry (e.g., real estate), rating (e.g., AA), remaining maturity (e.g., 54D), buy / sell direction, and trading volume information.

[0077] When S500 receives a price inquiry request from a trader for the aforementioned bond product, it determines the target counterparty matching the bond product through a multi-dimensional matching model based on the aforementioned counterparty matrix, outputs the target counterparty information, and / or sends a price inquiry instruction to the target counterparty.

[0078] In this embodiment, the candidate trading counterparties are preferably the traders' friends in the address book of associated chat tools such as Ideal. In this case, the constructed trading counterparty matrix is ​​the friend matrix.

[0079] See Figure 1 As shown, the counterparty (friend) recommendation scheme provided by this invention can mainly include three parts: constructing multi-dimensional profile tags, receiving and parsing investment instructions, and multi-dimensional matching calculation.

[0080] Before constructing multi-dimensional profile tags, multimodal data is first acquired. In this embodiment, the multimodal data may include structured data and unstructured data.

[0081] The structured data may specifically include historical transaction data and bond data using a preset standard data format. The bond data is the counterparty's bond holdings / demand data, including transaction element information such as bond (bond code), counterparty, remaining maturity, coupon rate, and transaction direction.

[0082] The unstructured data includes unstructured interaction data and unstructured real-time market data. The unstructured interaction data includes chat data from associated chat tools—such as chat information from chat tools like Ideal and / or Qtrade—specifically, it can be the NLP (Natural Language Processing) parsing results of chat logs. The unstructured real-time market data includes intended price data from associated chat tools—such as intended price data from the Ideal chat tool.

[0083] Then, multimodal data fusion is performed, and multi-dimensional profile labels are configured for each candidate trading counterparty based on the fused data.

[0084] Specifically, when the tags include precise tags, preference tags, and behavioral feature tags, the tag configuration method is as follows: For each candidate trading counterparty (such as a friend), after integrating the friend's real-time chat data and quotation data in the associated chat tool, perform data semantic analysis (which can be done using NLP parsing). Based on the semantic analysis results, determine whether the friend has a bond code that is expected to be traded. If so, use the bond code as the precise tag of the candidate trading counterparty and record the date of the data occurrence.

[0085] Furthermore, after integrating the friend's historical transaction data and bond data, semantic analysis is performed (which can be done using NLP parsing). Based on the semantic analysis results, it can be determined whether the friend has a preferred bond type—for example, friend A's transaction data may contain multiple ABS bonds. Sometimes, the bond type can be used as a preference tag for the friend—for example, for friend A mentioned above, an ABS bond tag can be configured, and the data occurrence date can be recorded.

[0086] Furthermore, after integrating the friend's historical transaction and pricing data, semantic analysis is performed (which can be done using NLP parsing). Based on the semantic analysis results, it is determined whether the friend has a preferred bond category—for example, friend B likes to buy commercial gold bonds, and friend C likes to buy bonds from specific regions. Sometimes, the bond category is used as a behavioral characteristic label for the friend, and the date the data occurred is recorded.

[0087] After configuring the multi-dimensional profile tags for each friend, construct the aforementioned friend matrix based on the friend's name and / or ID, multi-dimensional profile tags, data occurrence date, and data type. See also Figure 2The example illustrates the specific data format of a friend matrix. Elements in the matrix include at least the name and / or ID of the candidate trading counterparty, precise tags, preference tags, behavioral characteristic tags, data occurrence date, and data type information. Specifically, the data type can be chat-related, market data-related, or transaction-related.

[0088] Using the above approach, the system can tag friends with multiple dimensions based on chat data (NLP analysis), transaction data, coupon data, market data, and other data.

[0089] Upon receiving an investment instruction, the instruction is parsed. For example, the investment manager's instruction is parsed to extract the trading characteristics of the bonds in the instruction, such as "bid 21 **Company MTN005 2000W", where "bid" indicates a buy and "**Company" indicates the issuer.

[0090] Based on the bond codes in the investment instructions, a multi-dimensional matching calculation is performed. Specifically, a matching score is calculated to determine the degree of match between candidate counterparties and bond products.

[0091] In this embodiment, the multi-dimensional matching model is configured to use a hierarchical matching method combining precise matching and association rule matching to calculate the matching score between bond products and candidate counterparties. First, precise matching is performed between the bond product's bond code and precise tags to obtain matched counterparties recommended to the trader (client). When the number of precisely matched counterparties does not meet a preset requirement, association rule matching of the tags is triggered to obtain more counterparties matching the aforementioned bond product and recommend them to the trader.

[0092] In practice, the preferred steps for calculating the matching score between bond products and candidate counterparties using a tiered matching method are as follows.

[0093] S510, for each bond product in the investment instruction, the counterparty matrix is ​​queried based on the bond product's bond code to obtain precise tags matching that bond code. The candidate counterparty information to which these precise tags belong is extracted to form the counterparty information to be matched. For each counterparty to be matched, the weights of the counterparty's precise tag data are adjusted according to a time-related tag factor parameter. Based on the adjusted weight values, a weighted algorithm is used to calculate the matching degree score between the counterparty and the bond product. All counterparties to be matched are sorted from largest to smallest matching degree scores, thus completing the precise matching.

[0094] S520, determine whether the number of counterparties to be matched has reached a preset threshold N, where N is an integer greater than or equal to 2. If the preset threshold N is reached, proceed to step S530. If the preset threshold N is not reached, proceed to step S540.

[0095] S530 outputs the top N sorted counterparties to be matched as target counterparty recommendations, and then ends.

[0096] S540, trigger association rule matching. At this time, first construct a counterparty-bond dataset based on the precise label and fuzzy label data in the counterparty matrix, then use a preset confidence algorithm to obtain the associated bond products that have the highest correlation with the aforementioned bond products and appear in the precise label data. Perform the aforementioned precise matching using the bond codes of the associated bond products to obtain the target counterparties related to the associated bond products, and output the target counterparties of the associated bond products as the target counterparties of the aforementioned bond products.

[0097] As an example, not a limitation, let's take a trader (e.g., Zhang San) who wants to buy / sell a bond product called Bond G. First, precise matching of tags is performed using the Bond G code to obtain recommended friends. When the number of precisely matched recommended friends does not meet the preset requirement N, the tag association rule matching is triggered. At this time, a friend-bond dataset can be constructed based on the precise tags, preference tags, and behavioral feature tags in the friend matrix. Then, a preset confidence algorithm is used to obtain the associated bond product Bond A that has the highest correlation with the aforementioned Bond G and appears in the precise tag data (friends who have traded Bond A have the highest probability of buying or selling Bond G). Precise matching is performed using the bond code of Bond A to obtain recommended friends that match Bond A. The recommended friends of Bond A are then output as the recommended friends of the aforementioned Bond G.

[0098] In this embodiment, the time-related labeling factor may specifically include a time enhancement factor and a time decay factor. The weight value corresponding to the time enhancement factor increases over time within a preset time range. The weight value corresponding to the time decay factor decreases over time within a preset time range.

[0099] Specifically, for bonds that have been purchased, because the willingness to sell these bonds increases over time, the weight of their labeling factor is configured to increase over time for bonds purchased within a preset time frame—e.g., 90 days. For bond price inquiry data, the probability of buying or selling decreases over time; therefore, the weight of its labeling factor is configured to decrease over time. For example, see [link to relevant documentation]. Figure 3The enhancement / deterioration factor configured in the system divides the time range corresponding to the preset time range into three stages: 0-7 days, 8-30 days, and 31-90 days. Starting from the date the data occurred, the weight values ​​of the enhancement / deterioration factor are configured differently for each stage as time goes on.

[0100] In this embodiment, when calculating the matching score, each data type is configured as an item based on the data type information in the counterparty matrix, and a maximum score max_score is configured for each data type item. The purpose of setting the maximum score max_score is to prevent the matching score from being fully occupied by a certain data type.

[0101] Preferably, the steps for calculating the matching score between the counterparty to be matched and the bond product using a weighted algorithm are as follows:

[0102] S511, for each data type item, calculate the individual label score (label_score) for each precise label, as follows: After determining the weight value corresponding to the label factor of the precise label based on the data occurrence date, multiply the weight value by the preset initial label score of the precise label to obtain the individual label score (label_score) of the precise label, i.e., label_score = label factor weight value * initial label score. Preferably, the initial label score is set by system default, for example, the system defaults to an initial label score of 1 for all precise labels.

[0103] S512, sum the tag scores of multiple precise tags belonging to the same data type item to obtain the score item_score for each item. The value of item_score does not exceed the maximum score max_score of the data type item configuration item, that is, item_score = Min( (max_score). The sum of tag scores for multiple precise tags. This is an example, not a limitation. Figure 3 The maximum score (max_score) for chat-related data is 50, for market data it is 30, and for transaction data it is 20.

[0104] S513, sum the scores of multiple data type items related to the bond product to obtain the total bond score bond_score. The value of bond_score does not exceed the pre-configured score threshold MAX, that is, bond_score = Min( (MAX). This is the sum of scores for multiple data type items.

[0105] As a typical example, in this embodiment, the score threshold MAX is set to 100 points, that is, bond_score = Min( ,100).

[0106] In this embodiment, the confidence algorithm is preferably the FP-Growth algorithm. Then, the steps for obtaining the associated bond products with the highest correlation to the aforementioned bond products and appearing in the precise labeling data can be as follows:

[0107] S521, Construct a counterparty-bond dataset based on the precise and fuzzy label data in the counterparty matrix. The counterparty-bond dataset records candidate counterparties and historical bond data purchased by the candidate counterparties.

[0108] For example, see Figure 4 As shown, an example of a completed counterparty (friend)-bond dataset is provided. This dataset records information about friends and the bonds they have purchased, including bonds A, B, C, D, E, F, and G. (See also...) Figure 4 listed.

[0109] S522, calculate the support of each bond based on the frequency of the bonds appearing in the counterparty-bond dataset; filter out bonds with a support lower than the preset minimum support based on the preset minimum support.

[0110] For example, if the preset minimum support is 2, bonds with a support less than 2 are filtered out. Bonds H, I, J, L, M, N, O, and P all appear less than twice.

[0111] S523, sort the remaining bonds after filtering according to their support to obtain the sorted dataset.

[0112] As an example, see the support levels (corresponding frequency of occurrence) of the remaining bonds after filtering. Figure 5 As shown. See the sorted dataset. Figure 6 As shown, those with high support are listed first, and those with low support are listed last.

[0113] S524, construct a frequent item tree (FP tree) based on the aforementioned sorted dataset.

[0114] See Figure 7 As shown, it is based on Figure 6 Constructed bond FP tree.

[0115] S525, based on the constructed FP treee, obtains the conditional pattern base data of the bond products in the investment instruction.

[0116] See Figure 8 As shown, this example illustrates the conditional pattern of bond product G that a trader wants to buy or sell.

[0117] S526, Calculate the confidence level of relevant bonds based on the conditional model base data of bond products and the support tables of each bond.

[0118] according to Figure 8 The confidence levels of the relevant bonds are calculated using the support table, as follows:

[0119] The confidence level of bond A -> bond G is 5 / 8 = 62.5%.

[0120] The confidence level of bond C -> bond G is 5 / 8 = 62.5%.

[0121] The confidence level of bond E to bond G is 4 / 8 = 50%.

[0122] S527, sort the bonds according to their confidence level, and select the bonds with the highest confidence level that appear in the precise label data as related bond products; perform precise matching on the related bond products to obtain target counterparty information.

[0123] Select Bond A, which has the highest confidence level and appears in the precise tag, as the associated bond product with the highest confidence level. Perform precise matching on Bond A using the precise matching algorithm to obtain the matching friends and matching degree scores of Bond A. Sort the matching friends of Bond A in descending order of matching degree scores, and use the sorted matching friend data as the recommended friends output for Bond G.

[0124] The solution provided by this invention constructs a friend matrix based on multimodal data and introduces enhancement and decay factors in multi-dimensional matching calculations. This addresses the problem of weakening purchase intent over time in price inquiry data, thereby improving the accuracy of friend recommendations. Furthermore, the hierarchical matching algorithm reduces model illusions and improves both efficiency and accuracy.

[0125] Another embodiment of the present invention provides a counterparty recommendation system for an electronic trading platform. The system includes a data acquisition module, a tag configuration module, a matrix construction module, an investment instruction parsing module, and a counterparty matching module.

[0126] The data acquisition module is used to acquire multimodal data, which is multi-source data related to bond trading obtained from different channels.

[0127] The tag configuration module is used to perform data fusion on the acquired multimodal data and configure multi-dimensional profile tags for each candidate counterparty based on the fused data. The multi-dimensional profile tags include precise tags and fuzzy tags. The precise tags are used to identify the bond codes that the candidate counterparty hopes to trade, and the fuzzy tags are used to identify the bond preference characteristics that the candidate counterparty hopes to trade.

[0128] The matrix construction module is used to construct a counterparty matrix that includes the tag information of all candidate counterparties based on the multi-dimensional profile tags of each candidate counterparty.

[0129] The investment instruction parsing module is used to parse the investment instruction upon receiving it and obtain the trading characteristic elements of the bond product to be traded. The trading characteristic elements include at least the bond code.

[0130] The counterparty matching module is used to collect the inquiry request initiated by the trader for the aforementioned bond product, determine the target counterparty matching the bond product through a multi-dimensional matching model based on the aforementioned counterparty matrix, output the target counterparty information, and / or send an inquiry instruction to the target counterparty.

[0131] Other technical features are described in the preceding embodiments and will not be repeated here.

[0132] In another embodiment of the present invention, a computer-readable storage medium is provided for storing a computer program executable by a processing unit, wherein the computer program, when executed by the processing unit, implements the method described above.

[0133] The storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] Other technical features are described in the preceding embodiments and will not be repeated here.

[0135] In the above description, the disclosure of this invention is not intended to limit itself to these aspects. Rather, within the scope of the objectives of this disclosure, components can be selectively and operationally combined in any number. Furthermore, terms such as “comprising,” “encompassing,” and “having” should be interpreted by default as inclusive or open-ended, rather than exclusive or closed, unless explicitly defined as such. All technical, scientific, or other terms are to be understood by those skilled in the art, unless defined as such. Public terms found in dictionaries should not be interpreted in the context of the relevant technical documents in an overly idealistic or impractical manner, unless explicitly defined as such in this disclosure. Any modifications or alterations made by those skilled in the art based on the foregoing disclosure are within the scope of the claims.

Claims

1. A counterparty recommendation method based on multimodal data, characterized in that... Including the following steps: Acquire multimodal data, which is multi-source data related to bond trading obtained from different channels; The acquired multimodal data is fused, and multi-dimensional profile labels are configured for each candidate counterparty based on the fused data. The multi-dimensional profile labels include precise labels and fuzzy labels. The precise labels are used to identify the bond codes that the candidate counterparty expects to trade, and the fuzzy labels are used to identify the bond preference characteristics that the candidate counterparty expects to trade. Based on the multi-dimensional profile tags of each candidate counterparty, a counterparty matrix including tag information of all candidate counterparties is constructed; Upon receiving an investment instruction, the system parses the instruction to obtain the transaction feature elements of the bond product to be traded, which include at least the bond code. When a trader initiates a price inquiry request for the aforementioned bond product, based on the aforementioned counterparty matrix, a multi-dimensional matching model is used to determine the target counterparty matching the bond product. The target counterparty information is then output and / or a price inquiry instruction is sent to the target counterparty. Specifically, a matching score is calculated to obtain the matching degree between the candidate counterparty and the bond product. The multi-dimensional matching model is configured to use a hierarchical matching method combining precise matching and association rule matching to calculate the matching degree score between the bond product and the candidate counterparty. The steps are as follows: S510, for each bond product in the investment instruction, the counterparty matrix is ​​queried based on the bond code of the bond product to obtain the precise label matching the bond code, and the candidate counterparty information to which these precise labels belong is extracted to form the counterparty information to be matched; for each counterparty to be matched, the weights of the precise label data of the counterparty to be matched are adjusted according to the time-related label factor parameters, and the matching degree score between the counterparty to be matched and the bond product is calculated by a weighted algorithm based on the adjusted weight values; the matching degree scores of all counterparties to be matched are sorted from largest to smallest; the time-related label factors include time enhancement factors and time decay factors; the weight value corresponding to the time enhancement factor increases over time within a preset time range; the weight value corresponding to the time decay factor decreases over time within a preset time range; S520, determine whether the number of counterparties to be matched has reached a preset number threshold N, where N is an integer greater than or equal to 2; if it is determined that the preset number threshold N has been reached, proceed to step S530; if it is determined that the preset number threshold N has not been reached, proceed to step S540. S530: Output the top N sorted counterparties to be matched as target counterparty recommendations, then end; S540, trigger association rule matching; at this time, after constructing the counterparty-bond dataset based on the precise label and fuzzy label data in the counterparty matrix, the associated bond product with the highest correlation with the aforementioned bond product and appearing in the precise label data is obtained through the preset confidence algorithm. The bond code of the associated bond product is used for precise matching to obtain the target counterparty related to the associated bond product. The target counterparty of the associated bond product is recommended as the target counterparty of the aforementioned bond product.

2. The method according to claim 1, characterized in that, The candidate trading counterparties are the traders' friends in the address book of the associated chat tool. In this case, the constructed trading counterparty matrix is ​​a friend matrix.

3. The method according to claim 1, characterized in that, The fuzzy label includes a preference label and a behavioral feature label. The preference label is used to identify the type of bonds that the candidate counterparty prefers during the transaction, and the behavioral feature label is used to identify the major categories of bonds that the candidate counterparty prefers during the transaction. The bond type is distinguished by the sub-type of the bond issuer, and the bond category is distinguished by the parent type of the bond issuer, with one or more sub-types under the parent type.

4. The method according to claim 3, characterized in that, The multimodal data includes structured data and unstructured data; The structured data includes historical transaction data and bond data using a preset standard data format; The unstructured data includes unstructured interactive data and unstructured real-time market data. The unstructured interactive data includes chat data from associated chat tools, and the unstructured real-time market data includes intended price quotes from associated chat tools.

5. The method according to claim 4, characterized in that, The steps for configuring multi-dimensional profile tags for each candidate counterparty based on fused data include: For each candidate counterparty, after integrating the real-time chat data and quotation data of the candidate counterparty in the associated chat tool, data semantic analysis is performed. Based on the semantic analysis results, it is determined whether the candidate counterparty has the bond code of the desired transaction. If so, the bond code is used as the precise label of the candidate counterparty, and the date of data occurrence is recorded. Furthermore, after integrating the historical transaction data and bond data of the candidate counterparty, data semantic analysis is performed. Based on the semantic analysis results, it is determined whether the candidate counterparty has a preferred bond type. If so, the bond type is used as the preference label for the candidate counterparty, and the date of data occurrence is recorded. Furthermore, after integrating the historical transaction data and quotation data of the candidate counterparty, data semantic analysis is performed. Based on the semantic analysis results, it is determined whether the candidate counterparty has a preferred bond category. If so, the bond category is used as a behavioral characteristic label for the candidate counterparty, and the data occurrence date is recorded.

6. The method according to claim 5, characterized in that, After configuring the multi-dimensional profile tags for each candidate counterparty, construct the aforementioned counterparty matrix based on the candidate counterparty's name and / or number, multi-dimensional profile tags, data occurrence date, and data type. The elements of the counterparty matrix include at least the name and / or number of the candidate counterparty, precise tags, preference tags, behavioral characteristic tags, data occurrence date, and data type information.

7. The method according to any one of claims 1-6, characterized in that, Based on the data type information in the counterparty matrix, each data type is configured as an item, and a maximum score (max_score) is assigned to each data type item. The steps for calculating the matching score between the counterparty and the bond product using a weighted algorithm are as follows: For each data type item, calculate the individual label score (label_score) for each precise label, as follows: After determining the weight value corresponding to the tag factor of the precise tag based on the date of data occurrence, the weight value is multiplied by the preset initial tag score of the precise tag to obtain the single tag score of the precise tag, namely, tag_score = tag factor weight value * tag initial score; The tag scores of multiple precise tags belonging to the same data type item are summed to obtain the score item_score for each item. The value of item_score does not exceed the maximum score max_score of the data type item configuration item, that is, item_score = Min( (max_score); The scores of multiple data type items related to the bond product are summed to obtain the total bond score, bond_score. The value of bond_score does not exceed a pre-configured score threshold MAX, i.e., bond_score = Min( ,MAX).

8. The method according to claim 1, characterized in that, The confidence algorithm is the FP-Growth algorithm. The steps for obtaining the most relevant bond products that appear in the precise labeling data are as follows: A counterparty-bond dataset is constructed based on the precise and fuzzy label data in the counterparty matrix. The counterparty-bond dataset records candidate counterparties and historical bond data purchased by the candidate counterparties. Calculate the support level for each bond based on the frequency of its appearance in the counterparty-bond dataset; Based on a preset minimum support, bonds with support values ​​less than the minimum support are filtered out. The remaining bonds after filtering are sorted by support to obtain the sorted dataset; Construct a frequent item tree (FP tree) based on the sorted dataset mentioned above. Based on the constructed FP tree, the conditional pattern base data of the bond products in the investment instructions are obtained; Based on the conditional model base data of bond products, and combined with the support tables of each bond, the confidence level of the relevant bonds is calculated. The bonds are sorted by their confidence level, and the bonds with the highest confidence level that appear in the precise labeling data are identified as related bond products. The relevant bond products are precisely matched to obtain information on the target counterparty.

9. A counterparty recommendation system for an electronic trading platform according to the method of claim 1, characterized in that... include: The data acquisition module is used to acquire multimodal data, which is multi-source data related to bond trading acquired from different channels; The tag configuration module is used to perform data fusion on the acquired multimodal data and configure multi-dimensional profile tags for each candidate counterparty based on the fused data. The multi-dimensional profile tags include precise tags and fuzzy tags. The precise tags are used to identify the bond codes that the candidate counterparty hopes to trade, and the fuzzy tags are used to identify the bond preference characteristics that the candidate counterparty hopes to trade. The matrix building module is used to construct a counterparty matrix that includes the tag information of all candidate counterparties based on the multi-dimensional profile tags of each candidate counterparty; The investment instruction parsing module is used to parse the investment instruction upon receiving it and obtain the transaction feature elements of the bond product to be traded. The transaction feature elements include at least the bond code. The counterparty matching module is used to collect the inquiry requests initiated by traders for the aforementioned bond products, and after determining the target counterparty that matches the bond products through a multi-dimensional matching model based on the aforementioned counterparty matrix, output the target counterparty and / or send an inquiry instruction to the target counterparty.

Images