AI business resource intelligent matching method based on multi-dimensional data

Abstract

The invention relates to the technical field of AI business resource matching, and discloses an AI business resource intelligent matching method based on multi-dimensional data. The method comprises the following steps: constructing dynamic portraits of a resource demander and a resource supplier; performing multi-round iterative depth feature extraction on the dynamic portrait to obtain a core demand and supply feature vector; projecting the feature vectors to a unified feature space to calculate the proximity, and forming a potential matching pair set; performing matching feasibility evaluation and screening to obtain an optimized matching pair; and finally applying a dynamic matching constraint to generate a matching decision. According to the method, through deep iteration feature extraction and unified space projection, more accurate description of user intentions and discovery of deep semantic association are realized, and the matching accuracy and breadth are improved; and in combination with postposition dynamic constraint optimization, the commercial applicability and the performability of a matching result are enhanced.

Description

Technical Field

[0001] This invention relates to the field of AI business resource matching technology, specifically to an AI business resource intelligent matching method based on multi-dimensional data. Background Technology

[0002] In the current field of business resource matching, systems based on keywords or static tags are commonly used to associate supply and demand. These systems typically rely on pre-defined fixed attributes of resource demanders and suppliers, such as industry classification, product specifications, and geographical location, and use rule engines or simple similarity calculation models for filtering and recommendation. Existing technical solutions are effective in processing static, standardized information, but they are difficult to adapt to the highly dynamic and complex real-world scenarios in business activities.

[0003] Existing technologies have shortcomings. Static tagging systems cannot effectively capture and reflect the behavioral changes and intention shifts of both supply and demand sides over time. Historical data and real-time interaction data are often processed in isolation, lacking an effective integration mechanism, resulting in rigid user profiles that fail to accurately depict their current core needs and supply capabilities. Matching models based on fixed rules lack flexibility, struggle to handle unstructured data, and cannot uncover deeper semantic relationships beyond surface tags, easily missing potential high-quality matching opportunities.

[0004] After initially generating matching pairs, traditional constraint processing methods typically involve setting hard filtering conditions at the beginning of the matching process. While this rigid screening mechanism ensures compliance with basic conditions, it may prematurely exclude potential matching solutions that, although not fully meeting the initial constraints, could become highly commercially valuable after some coordination or optimization. The system lacks the ability to introduce dynamic, flexible constraints for re-optimization in the final decision-making stage, limiting the overall quality and commercial viability of the matching results. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based intelligent matching method for business resources based on multi-dimensional data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides an AI-based intelligent matching method for business resources based on multi-dimensional data, the method comprising: A dynamic profile of resource demanders and resource suppliers is constructed during the business resource matching process. The dynamic profile is generated by fusing historical interaction records and real-time behavioral data. The dynamic profile is subjected to multiple rounds of iterative deep feature extraction to obtain the core demand feature vector of the resource demander and the core supply feature vector of the resource supplier; Project the core demand feature vector and the core supply feature vector onto a unified feature space, and calculate the proximity of the core demand feature vector and the core supply feature vector in the unified feature space. Based on the proximity, resource demanders and resource suppliers are initially associated to form a set of potential matching pairs. For each potential matching pair in the set of potential matching pairs, a matching feasibility assessment is performed to generate a matching feasibility score; The potential matching pair set is filtered based on the matching feasibility score to obtain an optimized matching pair set; Dynamic matching constraints are applied to the optimized matching set to generate the final matching decision result.

[0007] Preferably, the step of constructing the dynamic profile further includes: Collect historical inquiry data, historical transaction data, and real-time browsing behavior sequences from resource demanders; Collect historical price quotes, historical contract fulfillment data, and real-time inventory change sequences from resource suppliers; Feature fusion technology is used to integrate the historical inquiry data, the historical transaction data, and the real-time browsing behavior sequence into a dynamic profile of the resource demander; Feature fusion technology is used to integrate the historical price data, the historical performance data, and the real-time inventory change sequence into a dynamic profile of the resource supplier.

[0008] Preferably, the multi-round iterative deep feature extraction step further includes: The dynamic profile of the resource demander is input into a deep neural network model for the first round of feature transformation to obtain the primary demand features of the resource demander. The dynamic profile of the resource supplier is input into a deep neural network model for the first round of feature transformation to obtain the primary supply features of the resource supplier. The primary demand features are compared with the demand features in historical successful matching cases, and the differences are extracted as the secondary demand features of the resource demanders. The primary supply characteristics are compared with the supply characteristics in historical successful matching cases, and the differences are extracted as the secondary supply characteristics of the resource supplier. The primary demand features and the secondary demand features are weighted and concatenated to form the core demand feature vector of the resource demander. The primary supply characteristics and the secondary supply characteristics are weighted and concatenated to form the core supply characteristic vector of the resource supplier.

[0009] Preferably, the step of calculating proximity further includes: Determine the metric criteria in the unified feature space; According to the metric criteria, calculate the cosine similarity metric between the core demand feature vector and the core supply feature vector; The cosine similarity metric is normalized and mapped to a numerical range between zero and one. This value is then used as the proximity between the core demand feature vector and the core supply feature vector in the unified feature space.

[0010] Preferably, the step of forming a set of potential matching pairs further includes: Calculate the proximity for all combinations of resource demanders and resource suppliers; Set a proximity threshold and filter out combinations of resource demanders and resource suppliers whose proximity is greater than the proximity threshold; The selected combinations are sorted from high to low according to their proximity. The set of potential matching pairs is formed by selecting a preset number of combinations that rank highly.

[0011] Preferably, the matching feasibility assessment step further includes: For each potential matching pair in the set of potential matching pairs, analyze the credit rating of the resource demander involved in the potential matching pair and the service level agreement of the resource supplier involved in the potential matching pair. Analyze the budget constraints of the potential match on the resource demanders and the capacity constraints of the potential match on the resource suppliers. Based on the credit rating, the service level agreement, the budget constraint, and the capacity constraint, the matching feasibility score of the potential matching pair is calculated by the rule engine.

[0012] Preferably, the step of obtaining the optimized set of matching pairs further includes: The matching feasibility score of each potential matching pair in the potential matching pair set is compared with a preset feasibility threshold; Potential matching pairs whose matching feasibility score is lower than the preset feasibility threshold are eliminated; The remaining potential matching pairs are sorted in descending order according to the matching feasibility score; Select a predetermined number of potential matching pairs that are ranked at the top to form the optimized matching pair set.

[0013] Preferably, the step of applying dynamic matching constraints further includes: Real-time monitoring of market environment fluctuation data, and adjustment of matching priority rules based on the market environment fluctuation data; Obtain the current system load data and adjust the upper limit of the number of concurrent matching tasks based on the system load data; The adjusted matching priority rules and the adjusted maximum number of concurrent matching tasks are used as the dynamic matching constraints. The dynamic matching constraints are applied to the optimized matching pair set, and the final matching decision result is output.

[0014] Preferably, applying the dynamic matching constraints to the optimized matching pair set and finally outputting the final matching decision result includes: Based on the matching priority rules in the dynamic matching constraints, calculate the current priority score for each optimized matching pair in the optimized matching pair set. Based on the upper limit of the number of concurrent matching tasks in the dynamic matching constraints, determine the threshold of the number of matching pairs that can be processed in the current period. The optimized matching pair set is reordered in descending order of current priority score; Select the top N optimized matching pairs from the reordered set of optimized matching pairs, where N equals the threshold number of matching pairs that can be processed in the current period; For each selected optimized matching pair, a decision record containing matching details is generated, and the decision records are summarized into the final matching decision result.

[0015] Preferably, the step of outputting the final matching decision result further includes: The optimized matching pair set is reordered according to the matching priority rules in the dynamic matching constraints; Based on the upper limit of the number of concurrent matching tasks in the dynamic matching constraints, a specified number of matching pairs are selected from the reordered optimized matching pair set; Generate a unique match identifier for each selected match pair; The matching pair information with the matching identifier is sent to the corresponding resource demander terminal and resource supplier terminal.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This method employs a multi-round iterative deep feature extraction mechanism to progressively refine the features of dynamic profiles that integrate historical interaction records and real-time behavioral data. Unlike simple feature extraction with a single forward propagation, this method uses an implicit feedback loop to recalibrate data weights in each round of feature extraction, gradually focusing on key information that best reflects the user's true intentions and core capabilities. This progressive mining approach effectively separates short-term fluctuation noise from long-term stable preferences, thereby generating more robust and representative core feature vectors. The continuous update mechanism of the dynamic profile ensures that the feature vectors synchronously reflect the user's latest state changes, providing a fundamental guarantee for improved matching accuracy.

[0017] By mapping the core feature vectors of demand and supply sides to a unified semantic space, this fundamentally changes the limitations of traditional keyword-based literal matching. This projection transformation aligns the semantic representations of heterogeneous data through a deep learning model, making demand and supply information from different sources and with different structures comparable within a unified metric space. Within this space, the geometric relationships between vectors directly reveal potential business fit, identifying deep connections beyond surface labels. The subsequently introduced dynamic constraints no longer serve as pre-screening but as subsequent optimization variables. This approach preserves the diversity of potentially high-quality matches while ensuring the feasibility of the final matching scheme through flexible constraint adjustments. The entire matching process thus achieves closed-loop optimization from semantic understanding to business decision-making, while simultaneously guaranteeing the accuracy and feasibility of the matching results. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based intelligent matching method for business resources based on multi-dimensional data as described in this invention. Figure 2 A flowchart for constructing a dynamic profile; Figure 3 A flowchart for forming a set of potential matching pairs; Figure 4 A comparison chart of priority scores for different types of business resources; Figure 5 The final priority score is used to construct an analytical histogram. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides an AI-based intelligent matching method for business resources based on multi-dimensional data. The method includes: constructing dynamic profiles of resource demanders and suppliers during the business resource matching process. These dynamic profiles are not static attribute sets but are generated by fusing historical interaction records and real-time behavioral data, dynamically reflecting the current state and intentions of both parties. The constructed dynamic profiles undergo multiple rounds of iterative deep feature extraction. This process aims to go beyond surface attributes and uncover intrinsic, stable feature representations to obtain the core demand feature vector of the resource demander and the core supply feature vector of the resource supplier. These two core feature vectors are projected onto a pre-learned unified feature space, where the proximity between them is calculated. This metric reflects the semantic matching degree between demand and supply.

[0021] Based on the calculated proximity, the system initially associates resource demanders and suppliers, selecting combinations with high proximity to form a potential matching pair set. Not all potential matching pairs can successfully translate into actual transactions; therefore, a matching feasibility assessment is required for each pair in the potential matching pair set. This assessment comprehensively considers practical constraints such as credit, service level, budget, and capacity, generating a quantified matching feasibility score. Based on this score, the potential matching pair set is further filtered, eliminating combinations with low feasibility to obtain an optimized matching pair set. Real-time factors such as market fluctuations and system load are applied as dynamic matching constraints to the optimized matching pair set. By adjusting priorities and concurrency levels, the final matching decision result is generated, thus completing the entire intelligent matching process.

[0022] Example 1: See Figure 2 In its implementation, the process of building a dynamic profile of resource demanders begins with data collection. The system collects historical interaction records from multiple business databases through pre-integrated data interfaces. These historical records include historical inquiry data and historical transaction data. Historical inquiry data records product or service inquiries made by resource demanders to different resource suppliers over a period of time, including the specifications, quantity, expected price range, and timestamp of the inquired item. Historical transaction data records the final transactions completed by resource demanders, including details of the transacted item, the final transaction price, counterparty information, and contract performance status. Simultaneously, the system captures real-time browsing behavior sequences of resource demanders through a real-time log collection module. These sequences include page browsing records on the business platform, duration of stay on product detail pages, search keyword sequences, and changes in favorites or watchlists. This historical and real-time data is then transmitted to a feature processing engine for subsequent fusion.

[0023] When integrating resource demander data using feature fusion technology, the system first cleans and standardizes historical inquiry data, historical transaction data, and real-time browsing behavior sequences. This process includes removing outliers, handling missing values, and converting unstructured log text into a structured data format. The cleaned data is then aligned according to a unified time window. The aligned data is then feature-mapped through a shared embedding layer, a trained fully connected neural network that transforms heterogeneous multi-source data into dense vector representations within the same vector space. Finally, the vectors mapped from historical inquiry data, historical transaction data, and real-time browsing behavior sequences are concatenated. The resulting composite vector is the dynamic profile of the resource demander, a numerical representation that comprehensively reflects both long-term preferences and short-term intentions.

[0024] In practice, the process of building a dynamic profile of resource suppliers is similar to that of resource demanders, but the data sources differ. The system collects historical interaction records from resource suppliers, specifically including historical quotation data and historical performance data. Historical quotation data contains quotation information provided by resource suppliers for historical inquiries, such as quoted price, quotation validity period, minimum order quantity, and additional service terms. Historical performance data comes from the supply chain management system and records the fulfillment status of completed orders by resource suppliers, including on-time delivery rate, product quality pass rate, after-sales service response time, and customer satisfaction rating. Simultaneously, the system obtains real-time inventory change sequences through the resource supplier's enterprise resource planning (ERP) system interface. These sequences dynamically reflect changes in the inventory levels of raw materials, semi-finished products, and finished products, inbound and outbound records, and safety stock early warning information.

[0025] When integrating resource supplier data using feature fusion technology, the feature processing engine also performs data preprocessing on historical pricing data, historical fulfillment data, and real-time inventory change sequences, including data cleaning, format standardization, and timestamp alignment. The aligned data is then fed into an embedding layer specifically designed for supplier data. This layer encodes the commercial terms information from historical pricing data, the service quality information from historical fulfillment data, and the supply capacity information from real-time inventory change sequences into a unified, high-dimensional numerical vector. During this process, both quantitative and qualitative indicators from the historical fulfillment data are quantified as numerical features for embedding calculations. The vector output from the embedding layer represents a dynamic profile of the resource supplier, comprehensively reflecting its pricing strategy, fulfillment reputation, and real-time supply capacity.

[0026] It is understandable that the first stage of the multi-round iterative deep feature extraction process is the primary feature transformation. The dynamic profile of the resource demander is fed into a deep neural network model. This model employs a multilayer perceptron structure, with its input layer dimension matching the dimension of the dynamic profile vector, while the output layer dimension is determined based on the design dimension of the core feature vector. Through the non-linear activation functions of its hidden layers, the deep neural network model learns the complex mapping relationship from the original dynamic profile to higher-level abstract features. The output of the deep neural network model is defined as the primary demand feature of the resource demander. Similarly, the dynamic profile of the resource supplier is input into a deep neural network model with the same structure but independent parameters. This model performs the same feature transformation operation, and its output is defined as the primary supply feature of the resource supplier. The primary demand and primary supply features are more generalized feature representations that are detached from the details of the original data.

[0027] In practice, the extraction of secondary features relies on comparison with historical successful matching cases. The system maintains a historical successful matching case library, which stores the feature vectors of demanders and suppliers who have successfully matched and ultimately reached satisfactory transactions. When extracting secondary demand features from resource demanders, the system performs a consistency comparison between the current primary demand features of the resource demander and several of the most similar successful demander features in the historical successful matching case library. This consistency comparison is achieved by calculating the cosine similarity and selecting the top K cases with the highest similarity. Subsequently, the difference vector between the current primary demand features and the mean of the demand features of these successful cases is calculated. This difference vector is defined as the secondary demand feature of the resource demander, capturing the degree of deviation and unique patterns between the current demand and the generally successful demand. The secondary supply features of resource suppliers are generated using the same process, namely, by comparing the differences between the primary supply features and the historical successful supplier features.

[0028] It can be understood that the core feature vector is formed by a weighted concatenation of primary and secondary features. For resource demanders, the system concatenates primary and secondary demand features to form a new long vector. A learnable weight allocation network dynamically calculates the relative importance weights of primary and secondary demand features in the final core demand feature vector based on the content of the current input dynamic profile. The weight coefficients output by the weight allocation network are used to weight and scale the portions of the concatenated vector corresponding to primary and secondary features. The weighted vector is then subjected to a linear transformation layer to adjust the dimensions, ultimately generating the core demand feature vector for resource demanders. The core supply feature vector for resource suppliers is generated using the exact same logic, obtained by dynamically weighting and transforming primary and secondary supply features. The core demand feature vector and the core supply feature vector are the final feature representations used for precise matching after multiple rounds of iterative deep refinement.

[0029] In practical implementation, training the deep neural network model is a crucial step. The parameters of both the deep neural network model used for primary feature transformation and the weight allocation network used for weight assignment need to be trained through supervised learning. Training data comes from accumulated historical matching records, including successfully matched positive samples and unmatched negative samples. The training objective is to optimize the loss function so that the model can learn to extract the feature representations most predictive of the final match success from the dynamic profile. The loss function is typically designed as a contrastive loss or triplet loss, aiming to shorten the distance between successfully matched demand and supply feature pairs while widening the distance between unmatched feature pairs. The model training process is performed periodically offline, using gradient descent and its variants to optimize the model parameters. The trained model parameters are then deployed to the online matching system for real-time feature extraction.

[0030] Optionally, the construction and updating of the historical successful matching case library needs to follow a specific strategy. Cases in the library should not only include successfully matched transaction records, but also feedback from both parties after the transaction and performance quality scores. Only successful matching cases that reach a certain satisfaction threshold will be included in the library as a comparison benchmark. The case library needs to be updated regularly, removing outdated cases and adding new successful cases to ensure that it reflects the current market environment and business trends. The update strategy for the case library can be based on a time window, such as retaining only cases from the last twelve months, or on dynamic evaluation, such as automatically removing cases when certain features differ significantly from current mainstream patterns. The quality of the case library directly affects the accuracy of secondary feature extraction, and thus the representativeness of the core feature vector.

[0031] Optionally, the embedding layer used in the feature fusion process can be pre-trained based on self-supervised learning. In practice, self-supervised learning paradigms such as autoencoders or masked language models can be used to pre-train the feature fusion embedding layer on large-scale unlabeled business data. The goal of pre-training is to enable the embedding layer to learn the inherent patterns and structures in the business data, such as the correlation between items and the regularity of user behavior sequences. The pre-trained embedding layer can obtain better initialization parameters, thereby converging faster in subsequent supervised model training and potentially improving the quality of the final generated feature vectors. The pre-training process is independent of the main matching model training and can be regarded as a preliminary preparation.

[0032] Example 2: See Figure 3In practice, projecting the core demand feature vector and the core supply feature vector to a unified feature space is the fundamental operation for calculating proximity. This unified feature space is obtained through training a dedicated deep metric learning network. The deep metric learning network uses a large number of historically successfully matched "resource demander-resource supplier" feature vector pairs as positive samples and randomly combined or failed-match feature vector pairs as negative samples. The training objective is to optimize the network parameters so that, in the unified feature space output by the deep metric learning network, the distance between feature vectors of positive sample pairs is as small as possible, while the distance between feature vectors of negative sample pairs is as large as possible. After training, the output of the last hidden layer of the deep metric learning network is defined as the unified feature space. The core demand feature vector and the core supply feature vector are forward-propagated through this network, and the resulting output vector is their projection into the unified feature space. The existence of the unified feature space makes feature vectors from different sources comparable, and its metric directly determines the matching bias.

[0033] Determining the metric in a unified feature space is a pre-defined configuration step; in this implementation, cosine similarity is explicitly used as the metric. The cosine similarity metric focuses on the directional differences between two feature vectors, while being insensitive to the absolute length of the vectors. This characteristic helps eliminate the bias caused by differences in feature vector magnitude due to variations in data volume. Based on the cosine similarity metric, the cosine similarity metric between the core demand feature vector and the core supply feature vector is calculated, and its mathematical expression is as follows: Where: symbol This represents a core demand feature vector for a specific resource demander, with the symbol... The core supply feature vector representing a specific resource supplier, operators The dot product operation represents the operation of two vectors, with the symbol... The Euclidean norm (modulus) of the core requirement feature vector is represented by the symbol. The Euclidean norm (modulus) of the core supply feature vector. Cosine similarity measure. The calculation results are within the range of interval Within the range, the closer the value is to 1, the more consistent the directions of the two vectors are, and the higher their similarity; the closer the value is to -1, the more opposite the directions are.

[0034] In practice, normalizing the cosine similarity metric is a necessary step to adapt it for subsequent threshold judgment. This is because the original range of the cosine similarity metric is... The proximity threshold set in subsequent steps is typically expected to be a value between 0 and 1, thus requiring a linear transformation. The mapping process uses the formula: in: Represents the proximity obtained after mapping. This represents the original cosine similarity metric value obtained from the calculation. After this transformation, the proximity... The numerical range is mapped to ,when When, proximity ;when When, proximity Ultimately, this proximity value between 0 and 1 will serve as a metric for the matching of the core demand feature vector and the core supply feature vector in a unified feature space.

[0035] The first step in forming a set of potential matching pairs is to calculate the proximity for all possible combinations of resource demanders and resource suppliers. Assuming the system currently has M resource demanders and N resource suppliers, the proximity of M multiplied by N combinations needs to be calculated. This is a computationally intensive task, and in practical deployments, distributed computing frameworks such as Spark or the parallel computing capabilities of GPUs are typically used to accelerate processing. The calculation process involves iterating through each resource demander, pairing its core demand feature vector with the core supply feature vector of each resource supplier in the system, and calculating the proximity for each pair according to the steps described above. The proximity calculation results for all combinations are stored in an M x N proximity matrix, where the row index corresponds to the resource demander identifier, the column index corresponds to the resource supplier identifier, and the matrix elements are the corresponding proximity values.

[0036] In practical implementation, setting a proximity threshold is a key parameter for controlling the quality of the potential matching pair set. The proximity threshold is a configurable floating-point number, and its value directly affects the number and quality of the selected combinations. A higher proximity threshold means stricter selection criteria, with only highly similar combinations being selected, potentially resulting in fewer matching pairs but higher accuracy. A lower proximity threshold will include more combinations with moderate similarity, possibly increasing matching opportunities but also introducing more noise. The specific value of the proximity threshold can be determined by analyzing historical data; for example, selecting a threshold that ensures a high proportion of combinations from historically successful matches are included. When the system performs the filtering, it iterates through each element in the proximity matrix, comparing the element value with the set proximity threshold, and selecting all "resource demander-resource supplier" combinations with a proximity greater than the threshold.

[0037] It's understandable that sorting the selected combinations by proximity from highest to lowest is to establish matching priorities. The sorting operation is based on the list of combinations obtained after filtering. Each item in the list contains a resource demander identifier, a resource supplier identifier, and a corresponding proximity value. The sorting algorithm arranges the list in descending order of proximity values, placing the combination with the highest proximity at the top of the list. This sorting structure provides an intuitive priority order for subsequent selection operations; the system can be confident that the combinations at the beginning of the list have a higher matching degree in the feature space. The sorted list is an ordered sequence of candidate matching pairs.

[0038] In some embodiments, when selecting a preset number of combinations that rank highly to form a potential matching pair set, the determination of the preset number needs to consider business requirements and system processing capacity. The preset number can be a fixed integer value, such as a maximum of 1000 potential matching pairs output per matching task; or it can be a dynamic value, such as a value set proportionally based on the total number of currently online resource demanders and resource suppliers. The system selects a specified number of combinations sequentially, starting from the beginning of the sorted list. These selected combinations are encapsulated into a set, namely the potential matching pair set. Each entry in the potential matching pair set contains the identifiers of resource demanders and resource suppliers with high matching degrees, along with their proximity. This set will serve as input for subsequent matching feasibility assessments.

[0039] Optionally, when calculating the proximity of large-scale combinations, approximate nearest neighbor search techniques can be used to improve efficiency. When the number of resource suppliers N is very large, calculating the exact proximity of each resource demander to all resource suppliers is costly. Approximate nearest neighbor search techniques, such as algorithms based on locality-sensitive hashing or hierarchical navigable small worlds, can quickly retrieve the Top-K candidates most similar to a specified resource demander from a massive number of resource suppliers, thus avoiding full computation. The system can construct an approximate nearest neighbor search index for the core supply feature vectors of all resource suppliers. When it is necessary to find potential matches for a resource demander, the system can quickly obtain a list of candidate suppliers with high proximity by querying this index, and then merge the query results of all resource demanders to form a global set of potential matching pairs.

[0040] Optionally, the proximity threshold can be designed to be dynamically adjusted rather than fixed. The system can monitor the conversion rate of the potential matching pair set in real time during the subsequent feasibility assessment and final transaction stages. If the conversion rate is found to be consistently low, it may indicate that the current proximity threshold is set too high, filtering out too many valuable potential matches. In this case, the proximity threshold can be appropriately lowered to expand the candidate set. Conversely, if the conversion rate is high but the candidate set is too large, causing a bottleneck in system processing, the proximity threshold can be appropriately increased to focus on the most crucial matching opportunities. This dynamic adjustment mechanism allows the system to adaptively optimize the efficiency and effectiveness of the matching process.

[0041] Example 3: In practical implementation, the matching feasibility assessment step conducts in-depth analysis of each potential matching pair in the potential matching pair set. The analysis process involves obtaining constraint data from multiple dimensions from external systems or internal databases. For a specific potential matching pair, such as a combination identified as "Demander A - Supplier B", the system first queries the enterprise credit information database to obtain the credit rating of resource demander A. The credit rating may be a level symbol or a numerical score, reflecting demander A's historical payment records, financial status, and business reputation. Simultaneously, the system retrieves the service level agreement text of resource supplier B from the contract management system. The service level agreement clearly defines supplier B's specific commitments regarding service quality, delivery time, after-sales support, and liability for breach of contract. The system extracts key indicators from the service level agreement text using natural language processing technology.

[0042] In practice, analyzing budget and capacity constraints is crucial for assessing the feasibility of actual transactions. The system extracts the budget constraint from resource demander A's current inquiry or procurement plan. This budget constraint typically includes the maximum acceptable price or total amount limit set for this purchase. Simultaneously, the system retrieves capacity constraint data from resource supplier B's enterprise resource planning (ERP) system or supply chain management platform. This capacity constraint data indicates the maximum quantity of products supplier B can produce or the maximum amount of services it can provide within a specific time period. For example, resource demander A's budget constraint might be "the purchase price per unit cannot exceed 1000 yuan," while resource supplier B's capacity constraint might be "the remaining capacity this month is 5000 units." The system needs to verify whether supplier B's price quote for the products required by demander A is within the budget constraint and whether demander A's purchase quantity is within supplier B's remaining capacity.

[0043] It is understandable that the calculation of the matching feasibility score based on credit rating, service level agreement, budget constraints, and capacity constraints is executed through a rules engine. The rules engine predefines a series of production rules, each corresponding to a specific business logic judgment condition. For example, one rule might be "IF credit rating >= 'A' THEN score += 20", and another rule might be "IF service level agreement includes 'on-time delivery rate >= 95%' THEN score += 15". The rules engine will trigger all rules that meet the conditions sequentially, accumulating the scores obtained after rule execution. Rules related to budget constraints and capacity constraints typically involve numerical comparisons, such as "IF quote <= budget limit THEN score += 25; ELSE score += 0" and "IF demand quantity <= remaining capacity THEN score += 30; ELSE score += 0". Ultimately, the sum of the scores contributed by all rules constitutes the initial matching feasibility score for the potential matching pair. This scoring process can be formally represented by the following formula: Where: symbol The symbol represents the calculated matching feasibility score. Represents the total number of rules configured in the rule engine, symbol Represents the weight of the score contributed after the k-th rule is triggered, symbol [symbol missing]. It is an indicator function that activates when the conditions set by the k-th rule are met by data such as the credit rating, service level agreement, budget constraints, and capacity constraints of the current potential matching pair. The value is 1 if the condition is true and 0 otherwise. The rule engine iterates through all rules and sums them up to output a matching feasibility score. .

[0044] In practice, obtaining the optimized set of matching pairs begins by comparing the matching feasibility score of each potential matching pair with a preset feasibility threshold. The preset feasibility threshold is a threshold value set by the system administrator based on historical transaction data analysis and business objectives, for example, 60 points. The system iterates through the set of potential matching pairs, comparing the matching feasibility score of each potential matching pair calculated by the rules engine with this preset feasibility threshold. For potential matching pairs with matching feasibility scores lower than the preset feasibility threshold, the system performs a removal operation, removing these matching pairs from the current candidate set.

[0045] It's understandable that sorting the remaining potential matches in descending order of their feasibility scores is to establish a processing priority. After removing matches below a preset feasibility threshold, the system is left with a subset where all matches have a feasibility score that meets the minimum requirement. The system then sorts this subset in descending order based on its feasibility scores, placing the potential matches with the highest feasibility scores at the beginning of the sequence and the potential matches with the lowest feasibility scores at the end. This arrangement ensures that highly feasible matches receive priority processing and high visibility.

[0046] In some embodiments, when selecting a preset number of potential matching pairs to form an optimized matching pair set, the strategy for determining the preset number may vary. The preset number can be an absolute value, such as 500 pairs; regardless of how many qualified matching pairs remain, only the top 500 pairs after sorting are selected. The preset number can also be a relative proportion, such as retaining only the top 70% of the matching pairs after sorting. According to the preset strategy, the system selects a specified number or proportion of potential matching pairs, starting from the top of the descending list. These selected matching pairs constitute the optimized matching pair set. Compared to the initial potential matching pair set, the optimized matching pair set not only guarantees similarity at the feature level but also possesses higher practical transaction feasibility, representing a set of high-quality matching candidates.

[0047] Optionally, the rules and their weights in the rules engine can be dynamically configured and updated. Business analysts can access the rules engine's management module through a graphical interface to add new rules, modify or deactivate existing rules, or adjust the weight of each rule without modifying the system code. This flexibility allows the matching feasibility assessment criteria to quickly adapt to changes in market rules, company policies, or regulatory requirements.

[0048] Optionally, for certain constraints involved in the feasibility scoring process, such as capacity constraints, flexibility in the time dimension can be considered. When evaluating capacity constraints, the system can not only check the remaining capacity at the current moment, but also examine the capacity forecast for a future time period. If resource demander A's demand date is relatively late, and resource supplier B has available capacity during that time period, even if current capacity is tight, the rule engine may still determine it as feasible and award a certain score.

[0049] Example 4: In practical implementation, real-time monitoring of market environment fluctuation data is a crucial input source for dynamically matching constraints. This data is obtained through external data interfaces. The system periodically pulls data from authorized financial market data providers, industry information platforms, and macroeconomic statistics agencies. This data includes, but is not limited to, price indices for specific raw materials, purchasing managers' indices reflecting overall industry demand, futures market price change rates related to business resources, and exchange rate fluctuations affecting cross-border trade. The system parses and formats the acquired raw market environment fluctuation data, extracts key indicators strongly correlated with the current type of business resource to be matched, and calculates the magnitude and trend of these indicators relative to the previous monitoring period.

[0050] In practice, adjusting matching priority rules based on market fluctuation data is an automated process based on predefined strategies. While matching priority rules might initially be primarily based on matching feasibility scores, the introduction of market fluctuation data adds dynamic adjustment factors. The system maintains a "market factor-rule adjustment" mapping table, which defines the priority adjustment strategies triggered by changes in different market indicators. For example, the mapping table might stipulate that when the price index of a major raw material rises by more than 5% week-on-week, the historical price stability weight of the supplier increases for all matching pairs involving that raw material; when the industry purchasing managers' index is below the boom-bust line for two consecutive months, priority is given to matching demanders with shorter payment terms and higher prepayment ratios. The system queries this mapping table based on real-time monitored market fluctuation data to determine the adjustment strategy that should take effect, thereby dynamically modifying or reweighting the calculation factors in the matching priority rules.

[0051] It's understandable that obtaining real-time system load data and adjusting the upper limit of concurrent matching tasks accordingly is crucial for ensuring stable system operation. System load data is collected from deployed server monitoring agents and task queue monitors, and specific metrics include CPU utilization, memory usage, the number of currently executing matching evaluation threads, the number of matching requests backlogged in the message queue, and the number of active database connections. The system sets safety thresholds for various metrics. When one or more system load data metrics approach or exceed these thresholds, the system triggers a protection mechanism, dynamically lowering the upper limit of concurrent matching tasks to prevent system overload and service unavailability. Conversely, when system load data metrics indicate sufficient resources, the system can appropriately increase the upper limit of concurrent matching tasks to improve overall matching throughput. This adjustment is real-time and flexible.

[0052] In practice, the adjusted matching priority rules and the adjusted maximum number of concurrent matching tasks are jointly defined as dynamic matching constraints. A dynamic matching constraint is a structured data object generated at the start of each matching decision cycle. The dynamic matching constraint object contains two core fields: one is the "effective priority rule set," which describes the complete rules and weights used to calculate the priority score of matching pairs in the current cycle; the other is the "maximum number of concurrent matches," which is the maximum number of final matching pairs the system is allowed to produce in the current cycle. The dynamic matching constraint object is loaded into the memory of the matching decision engine as a global configuration parameter for this round of decision-making.

[0053] Understandably, the first step in applying dynamic matching constraints to the optimized matching pair set is to calculate the current priority score based on the adjusted matching priority rules. Each optimized matching pair in the set originally had a matching feasibility score, but under the new dynamic matching constraints, a comprehensive current priority score needs to be calculated. The calculation process incorporates the impact of market environment fluctuations; for example, a formula for calculating the current priority score that considers supply stability might look like this: Where: symbol Represents the calculated current priority score, symbol This represents the matching feasibility score obtained in the previous steps for the optimized matching pair, with the symbol [symbol missing]. It is the weighting coefficient of the supply-side stability factor, symbol It is a stability score (e.g., on-time delivery rate) calculated based on the historical performance data of resource suppliers, with the symbol... It is the penalty coefficient for price index volatility, symbol [symbol missing]. This represents the absolute value of recent fluctuations in the relevant raw material price index. The system iterates through the set of optimized matching pairs and calculates a new current priority score for each pair using the formula described above or similar derived rules.

[0054] In practice, determining the threshold for the number of matching pairs that can be processed within the current period based on the upper limit of concurrent matching tasks in the dynamic matching constraint is a direct assignment operation. The value of the "Maximum Concurrent Matches" field in the dynamic matching constraint object is directly used as the threshold N for the number of matching pairs that can be processed within the current period. For example, if system monitoring shows that the current load is high, and the upper limit of the number of concurrent matching tasks is adjusted to 200, then the threshold N for the number of matching pairs that can be processed within the current period is equal to 200. This threshold N determines the final number of matching pairs that can be output.

[0055] The final filtering step involves reordering the optimized matching pair set from highest to lowest priority score and selecting the top N optimized matching pairs. The system uses the current priority score as the key field to sort the optimized matching pair set in descending order. In the sorted list, the optimized matching pair with the highest current priority score is placed at the top. Subsequently, the system selects N optimized matching pairs sequentially from the beginning of the list, where N equals the threshold for the number of matching pairs that can be processed within the current period. These N optimized matching pairs are the most worthwhile matching results to output under the current market environment and system state, after being filtered by dynamic constraints.

[0056] In practice, referring to Table 1, generating a decision record containing matching details for each selected optimal matching pair and summarizing it into the final matching decision result is the final output of the process. The system creates a structured decision record for each selected optimal matching pair. The decision record is typically stored in JavaScript object notation and includes information such as the matching pair's unique identifier, resource requester ID, resource supplier ID, matching feasibility score, current priority score, matching timestamp, and a summary of the main market environment factors and system load status that triggered the matching decision. All N decision records are collected into a list or array; this set is the final matching decision result. The final matching decision result is persisted to the database archive and simultaneously sent to the message bus to trigger subsequent notification or transaction processes.

[0057] Table 1: Mapping Table of Market Factors and Rule Adjustments Optionally, the collection of system load data and the adjustment of the maximum number of concurrent matching tasks can achieve more granular control. The system can set different maximum concurrency limits for different types of matching tasks. The system load monitor not only monitors the overall resource utilization but also the length and processing latency of various task queues. The "maximum number of concurrent matches" in the dynamic matching constraint can be a structure containing multiple upper limit values ​​for different task types, thereby achieving more refined resource allocation and load control, maximizing resource utilization while ensuring system response speed.

[0058] Optionally, the final matching decision can include a simple explanatory summary. In the decision record generated for each optimized matching pair, in addition to the original data, a text description can be automatically generated, briefly explaining why this matching pair was selected. For example, "This matching pair was prioritized because the demander has a high credit rating and the supplier can still provide stable quotes in the current environment of rising chip prices." Such explanatory summaries help business personnel understand the decision-making logic of artificial intelligence, increasing the credibility and acceptability of the results. Summary generation can utilize template filling technology, automatically combining key field values ​​from the decision record.

[0059] See Figure 4 This study presents the differences in priority scores for different business resource types before and after rule adjustments. Specifically, using "business resource type" as the classification dimension, it simultaneously displays the quantitative results of priority scores before (gray bars) and after (blue bars) rule adjustments. The score changes for different resource types reflect the actual impact of market factor rule adjustments: for example, the score for electronic components was 76.3 before the adjustment, increasing to 88.3 after, reflecting the effectiveness of priority weight adjustments triggered by corresponding market factors (such as fluctuations in core raw material prices); the agricultural product type shows an inverse change, with a score before the adjustment (81.4) higher than the score after (87.4), corresponding to differences in rule adjustment strategies for its associated market factors. At the parameter level, the score data for each resource type is calculated based on the priority rules in the dynamic matching constraints, directly reflecting the application effect of the "market factor-rule adjustment" mapping table in real-world scenarios.

[0060] Example 5: In specific implementation, the output of the final matching decision begins with reordering the optimized matching pair set according to the matching priority rules in the dynamic matching constraints. The dynamic matching constraints are a set of rules and parameters determined in the previous cycle based on the real-time market environment and system load. The matching priority rules may include multiple calculation factors, such as a weighted combination of matching feasibility score, supplier stability coefficient, and current market tightness index. The system traverses each optimized matching pair in the optimized matching pair set and calculates a priority value for each optimized matching pair for final sorting according to the calculation formula of the matching priority rules defined in the dynamic matching constraints. After calculation, the system calls the sorting algorithm, using the calculated priority value as the key field, to sort the entire optimized matching pair set in descending order, so that the optimized matching pair with the highest priority value is located at the beginning of the ordered set, while the optimized matching pairs with lower priority values ​​are located at the end of the ordered set. This reordering process ensures that the highest priority matching opportunity is processed first in subsequent selection steps.

[0061] In practice, a key operation is selecting a specified number of matching pairs from the reordered set of optimized matching pairs based on the maximum number of concurrent matching tasks specified in the dynamic matching constraints. The dynamic matching constraints explicitly define the maximum number of concurrent tasks the system can handle within the current matching cycle; this value is determined based on the system's real-time load capacity, for example, a maximum of 150. The system starts from the top of the reordered set of optimized matching pairs and reads them sequentially until the number of read pairs reaches the maximum specified in the dynamic matching constraints. These selected optimized matching pairs constitute the subset of matching results that will ultimately be used for actual recommendations and pushes in this cycle. If the total number of optimized matching pairs in the reordered set is less than the maximum number of concurrent matching tasks, all optimized matching pairs are selected.

[0062] It is understandable that generating a unique match identifier for each selected matching pair is fundamental for effective tracking and management. The system assigns a globally unique identifier to each selected optimized matching pair. This identifier is typically generated using a standard algorithm, such as time-based UUID version 1 or UUID version 4 with sufficient randomness. The process of generating the match identifier ensures that no identifier duplication occurs, even when processing a large number of matching pairs simultaneously in a distributed system environment. The generated match identifier is immediately bound to the metadata information of this optimized matching pair, including the timestamp of the match generation, the involved demand party number, supplier number, and matching priority value. As the unique credential for this matching event, the match identifier will be used throughout subsequent storage, query, and notification processes.

[0063] In practice, sending the matching pair information with matching identifiers to the corresponding resource requester and resource supplier terminals is the final step in completing the matching loop. The system constructs the message body to be sent, using a structured data format. The message body must contain the previously generated matching identifier and detailed information related to this matching pair. This detailed information typically includes basic information about the resource requester and resource supplier, a summary of the matched business resources, a summary of the matching score calculated by the system, and suggested next steps. The system asynchronously sends the constructed message body to the corresponding resource requester and resource supplier terminals in the matching pair by querying their registered message receiving addresses, message queue topics, or mobile device push tokens. The sending process needs to ensure reliable message delivery, typically achieved through a message middleware acknowledgment mechanism.

[0064] Optionally, when sending matching information to terminals, the message content and format can be adapted based on terminal type and user preferences. For resource requester terminals, if the terminal is a mobile application, the message content can be more concise and include a deep link directly to the in-app negotiation page. For resource provider terminals, if the terminal is a web management backend, the message content can be more detailed, including more data metrics for rapid decision-making. Before sending, the system will appropriately convert and trim the general message body according to the supported message formats and content preferences declared in the terminal's registration information, generating a version most suitable for the terminal's display and processing, thereby improving user experience and operational efficiency. This adaptive sending improves the actual reach and conversion rate of matching results.

[0065] Optionally, message sending success or failure requires status tracking and retrying. After sending a message to the resource requester and resource supplier terminals, the system listens for delivery receipts from the message middleware or terminal confirmation interface. If no successful receipt is received within a certain time, the system marks the sending task as failed and adds it to the retry queue. Retry strategies typically include exponential backoff, i.e., waiting a short time to retry after the first failure, waiting a longer time to retry after the second failure, until the maximum number of retries is reached. For messages that ultimately fail to be sent, the system logs the message and may trigger an alarm, requiring intervention from operations personnel. This mechanism ensures the reliability of matching result notifications and avoids the loss of matching information due to temporary network fluctuations or terminal service unavailability. The entire sending, confirmation, and retry process is automatically completed by the message processing framework.

[0066] It's understandable that the matching identifier plays a crucial indexing role in subsequent interactions. Once the resource requester or supplier receives the matching notification, all subsequent requests they send via application programming interfaces (APIs) or user interfaces regarding this match—whether for queries, confirmations, rejections, or further negotiations—must include this matching identifier in the request body. The system can use the matching identifier to quickly retrieve the complete context information of this match from persistent storage, including intermediate data from the matching calculation process, thereby providing users with a seamless interactive experience and supporting complex subsequent business processes.

[0067] See Figure 5The weighted score composition of each optimized matching pair is presented in the form of a stacked bar chart. The score is a weighted aggregation of the matching feasibility score (50%), the supplier stability coefficient (30%), and the market tightness index (20%). Specifically, the horizontal axis represents different optimized matching pairs, and the vertical axis represents the weighted score. The layers of each stacked bar correspond to the three scoring dimensions: the blue layer represents the matching feasibility score (the core influencing factor, with the highest weight), the purple layer represents the supplier stability coefficient, and the yellow layer represents the market tightness index. The data distribution shows that "Matching Pair 09" ranks first with a weighted score of 85.3, and its matching feasibility score (blue layer) is outstanding among all matching pairs. The scores of the other matching pairs decrease sequentially with the overall performance of the three dimensions, reflecting the weighted calculation logic of the priority rules in the dynamic matching constraints. This visually presents the priority score composition of each optimized matching pair, providing data visualization support for the final matching decision of "sorting and selecting matching pairs by priority," clearly reflecting the differences in the contributions of the three factors—matching feasibility, supplier stability, and market tightness—to the final priority.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent matching method for business resources based on multi-dimensional data, characterized in that, Includes the following steps: A dynamic profile of resource demanders and resource suppliers is constructed during the business resource matching process. The dynamic profile is generated by fusing historical interaction records and real-time behavioral data. The dynamic profile is subjected to multiple rounds of iterative deep feature extraction to obtain the core demand feature vector of the resource demander and the core supply feature vector of the resource supplier; Project the core demand feature vector and the core supply feature vector onto a unified feature space, and calculate the proximity of the core demand feature vector and the core supply feature vector in the unified feature space. Based on the proximity, resource demanders and resource suppliers are initially associated to form a set of potential matching pairs. For each potential matching pair in the set of potential matching pairs, a matching feasibility assessment is performed to generate a matching feasibility score; The potential matching pair set is filtered based on the matching feasibility score to obtain an optimized matching pair set; Dynamic matching constraints are applied to the optimized matching set to generate the final matching decision result.

2. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 1, characterized in that, The step of constructing the dynamic profile further includes: Collect historical inquiry data, historical transaction data, and real-time browsing behavior sequences from resource demanders; Collect historical price quotes, historical contract fulfillment data, and real-time inventory change sequences from resource suppliers; Feature fusion technology is used to integrate the historical inquiry data, the historical transaction data, and the real-time browsing behavior sequence into a dynamic profile of the resource demander; Feature fusion technology is used to integrate the historical price data, the historical performance data, and the real-time inventory change sequence into a dynamic profile of the resource supplier.

3. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 2, characterized in that, The multi-round iterative deep feature extraction step further includes: The dynamic profile of the resource demander is input into a deep neural network model for the first round of feature transformation to obtain the primary demand features of the resource demander. The dynamic profile of the resource supplier is input into a deep neural network model for the first round of feature transformation to obtain the primary supply features of the resource supplier. The primary demand features are compared with the demand features in historical successful matching cases, and the differences are extracted as the secondary demand features of the resource demanders. The primary supply characteristics are compared with the supply characteristics in historical successful matching cases, and the differences are extracted as the secondary supply characteristics of the resource supplier. The primary demand features and the secondary demand features are weighted and concatenated to form the core demand feature vector of the resource demander. The primary supply characteristics and the secondary supply characteristics are weighted and concatenated to form the core supply characteristic vector of the resource supplier.

4. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 3, characterized in that, The step of calculating the proximity further includes: Determine the metric criteria in the unified feature space; According to the metric criteria, calculate the cosine similarity metric between the core demand feature vector and the core supply feature vector; The cosine similarity metric is normalized and mapped to a numerical range between zero and one. This value is then used as the proximity between the core demand feature vector and the core supply feature vector in the unified feature space.

5. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 4, characterized in that, The step of forming a set of potential matching pairs further includes: Calculate the proximity for all combinations of resource demanders and resource suppliers; Set a proximity threshold and filter out combinations of resource demanders and resource suppliers whose proximity is greater than the proximity threshold; The selected combinations are sorted from high to low according to their proximity. The set of potential matching pairs is formed by selecting a preset number of combinations that rank highly.

6. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 5, characterized in that, The steps of the matching feasibility assessment further include: For each potential matching pair in the set of potential matching pairs, analyze the credit rating of the resource demander involved in the potential matching pair and the service level agreement of the resource supplier involved in the potential matching pair. Analyze the budget constraints of the potential match on the resource demanders and the capacity constraints of the potential match on the resource suppliers. Based on the credit rating, the service level agreement, the budget constraint, and the capacity constraint, the matching feasibility score of the potential matching pair is calculated by the rule engine.

7. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 6, characterized in that, The step of obtaining the optimized set of matching pairs further includes: The matching feasibility score of each potential matching pair in the potential matching pair set is compared with a preset feasibility threshold; Potential matching pairs whose matching feasibility score is lower than the preset feasibility threshold are eliminated; The remaining potential matching pairs are sorted in descending order according to the matching feasibility score; Select a predetermined number of potential matching pairs that are ranked at the top to form the optimized matching pair set.

8. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 7, characterized in that, The step of applying dynamic matching constraints further includes: Real-time monitoring of market environment fluctuation data, and adjustment of matching priority rules based on the market environment fluctuation data; Obtain the current system load data and adjust the upper limit of the number of concurrent matching tasks based on the system load data; The adjusted matching priority rules and the adjusted maximum number of concurrent matching tasks are used as the dynamic matching constraints. The dynamic matching constraints are applied to the optimized matching pair set, and the final matching decision result is output.

9. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 8, characterized in that, The step of applying the dynamic matching constraints to the optimized matching pair set and finally outputting the final matching decision result includes: Based on the matching priority rules in the dynamic matching constraints, calculate the current priority score for each optimized matching pair in the optimized matching pair set. Based on the upper limit of the number of concurrent matching tasks in the dynamic matching constraints, determine the threshold of the number of matching pairs that can be processed in the current period. The optimized matching pair set is reordered in descending order of current priority score; Select the top N optimized matching pairs from the reordered set of optimized matching pairs, where N equals the threshold number of matching pairs that can be processed in the current period; For each selected optimized matching pair, a decision record containing matching details is generated, and the decision records are summarized into the final matching decision result.

10. The AI-based intelligent matching method for business resources based on multi-dimensional data as described in claim 9, characterized in that, The step of outputting the final matching decision result further includes: The optimized matching pair set is reordered according to the matching priority rules in the dynamic matching constraints; Based on the upper limit of the number of concurrent matching tasks in the dynamic matching constraints, a specified number of matching pairs are selected from the reordered optimized matching pair set; Generate a unique match identifier for each selected match pair; The matching pair information with the matching identifier is sent to the corresponding resource demander terminal and resource supplier terminal.

Images