Supplier matching method and system based on multi-source heterogeneous data fusion

Abstract

The invention discloses a supplier matching method and system based on multi-source heterogeneous data fusion, and relates to the field of building engineering collection analysis, and the method comprises the steps: carrying out the feature analysis of a purchase demand text, and generating a structured purchase demand vector; dynamically adjusting the weight of the key index in the supplier portrait based on multi-source data fusion and historical time sequence prediction; selecting a feature mapping relation in combination with engineering types, extracting Top-K portrait indexes most relevant to demands, and performing weighted aggregation on key portrait indexes in the same supplier to generate exclusive representation vectors thereof; and finally, through similarity calculation of the vector and a purchase demand vector, high-precision candidate supplier screening is realized.

Description

Technical Field

[0001] The present application relates to the technical field of construction project procurement analysis, and in particular to a supplier matching method and system based on multi-source heterogeneous data fusion. Background Art

[0002] In the construction industry, supplier selection and management permeate the entire project bidding, procurement, and contract fulfillment process, directly impacting project quality, schedule, and cost control. With the widespread application of digitalization and big data technologies, a diverse and heterogeneous data set has emerged within the industry, including contract management systems, quality monitoring platforms, internal corporate financial data, and external data such as third-party credit ratings, legal proceedings, and public opinion monitoring. However, existing procurement systems lack the ability to integrate this data, making it difficult to fully leverage the value of this multi-source information in supplier management.

[0003] Most current systems still suffer from data silos and inadequate timeliness. For one thing, there's a lack of effective data connectivity and sharing mechanisms between internal systems and external platforms. Furthermore, external credit and risk information often lags in updates. When buyers make decisions, the data they rely on is often a static snapshot from yesterday or last month, which doesn't reflect the supplier's latest performance status or potential changes.

[0004] In addition, the construction of supplier portraits mostly relies on static information such as historical average fulfillment rate, qualification level and financial indicators, and key elements such as dynamic feedback and on-site evaluation are not updated in a timely manner; corresponding to this is the diversity and timeliness of procurement needs themselves - different project types have different focuses on quality, delivery timeliness, technical qualifications, etc., but existing matching models often use fixed weight configurations, ignoring the actual matching preferences brought about by differences in demand, resulting in generalized recommendation results and insufficient accuracy, requiring a large amount of manual secondary screening. Summary of the Invention

[0005] The present application provides a supplier matching method, system, storage medium, computer program product and electronic device based on multi-source heterogeneous data fusion, which is used to at least solve the problems of data silos, insufficient response time and rigid matching mode in the current related technologies.

[0006] In a first aspect, an embodiment of the present application provides a supplier matching method based on multi-source heterogeneous data fusion, the method comprising: performing feature parsing on an engineering procurement requirement text according to a preset set of procurement feature dimensions to generate a corresponding procurement requirement vector; the procurement feature dimensions include any one of the following: quality attention, fulfillment timeliness, risk tolerance, and price sensitivity; determining a monitoring real-time value of a supplier portrait indicator based on multiple data sources, and when the deviation between the monitoring real-time value of the supplier portrait indicator and the corresponding indicator prediction value exceeds a preset amplitude threshold, updating the supplier portrait by adjusting the portrait weight of the supplier portrait indicator; the indicator prediction value is a real-time prediction value calculated based on historical indicator data of the supplier portrait indicator, and the supplier portrait indicator includes any one of the following: fulfillment delay rate, quality pass rate, financial health index, and judicial public opinion compliance; selecting a feature mapping relationship according to the type of procurement project, extracting the Top-K feature subset with the highest correlation with the procurement requirement vector from the updated supplier portrait, and weighting it to generate a supplier-specific representation vector; calculating the similarity distance between the supplier-specific representation vector and the procurement requirement vector to screen candidate suppliers.

[0007] In the second aspect, the embodiment of the present application provides a supplier matching system based on multi-source heterogeneous data fusion, the system comprising: a procurement demand analysis unit, for performing feature analysis on the engineering procurement demand text according to a preset procurement feature dimension set to generate a corresponding procurement demand vector; the procurement feature dimension includes any one of the following: quality attention, fulfillment time, risk tolerance and price sensitivity; a supplier portrait update unit, for determining the monitoring real-time value of the supplier portrait indicator based on multiple data sources, and when the deviation between the monitoring real-time value of the supplier portrait indicator and the corresponding indicator predicted value exceeds a preset amplitude threshold, adjusting the supplier portrait indicator portrait weights are used to update the supplier portrait; the indicator prediction value is a real-time prediction value calculated based on the historical indicator data of the supplier portrait indicator, and the supplier portrait indicator includes any one of the following: fulfillment delay rate, quality pass rate, financial health index and judicial public opinion compliance; a Top-K vector generation unit is used to select feature mapping relationships according to the procurement project type, extract the Top-K feature subset with the highest correlation with the procurement demand vector from the updated supplier portrait, and weightedly generate a supplier-specific representation vector; a supplier screening unit is used to calculate the similarity distance between the supplier-specific representation vector and the procurement demand vector to screen candidate suppliers.

[0008] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the supplier matching method based on multi-source heterogeneous data fusion of any embodiment of the present application.

[0009] In a fourth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the supplier matching method based on multi-source heterogeneous data fusion of any embodiment of the present application are implemented.

[0010] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the supplier matching method based on multi-source heterogeneous data fusion of any embodiment of the present application.

[0011] The supplier matching method and system based on multi-source heterogeneous data fusion provided by this application can produce at least the following technical effects: (1) By setting multi-dimensional procurement feature dimensions (such as quality concern, fulfillment time, risk tolerance, and price sensitivity), we conduct structured analysis of engineering procurement requirements expressed in natural language and construct a vector modeling mechanism with the semantic features of procurement requirements as the core. This allows the demand expression to be transformed from unstructured subjective descriptions into quantifiable, multi-dimensional semantic vectors, significantly enhancing the system's ability to identify the differences between different engineering types and procurement goals.

[0012] (2) By collecting real-time indicator values ​​(such as fulfillment delay rate, quality qualification rate, financial health index, and judicial public opinion compliance) from multiple heterogeneous platforms and comparing them with the predicted values, a real-time monitoring and deviation control model is introduced into the dynamic update mechanism of the supplier portrait. When there is a significant deviation between the performance of the portrait indicator and the predicted value based on historical indicator data, a weight adjustment strategy can be triggered to achieve adaptive updating of the supplier portrait.

[0013] (3) By setting feature mapping relationships for different project types, the top-K feature subsets most relevant to current needs are extracted from the updated supplier profiles. Based on this, a supplier-specific representation vector is generated, and a feature mapping and selection mechanism is established between the demand side and the supply side. This effectively extracts the associated features between the profile dimensions and the procurement target. By introducing a similarity distance algorithm, accurate supplier matching and ranking decisions based on vector calculations are achieved. This improves the focus of feature matching, ensures that key capability dimensions are prioritized during the candidate screening process, and avoids the interference of weakly related or irrelevant features on the matching calculation.

[0014] Through this technical solution, the dynamic analysis of procurement demand feature vectors is integrated with multi-source heterogeneous data, achieving real-time updates and adaptive weight adjustments to supplier portraits, which can accurately reflect the latest status of suppliers in multiple dimensions such as fulfillment timeliness, quality qualification rate, financial health, and compliance risks. In addition, based on the matching of procurement project types with demand-side characteristics, it is possible to intelligently screen key supplier features that are highly relevant to actual needs, generate targeted supplier-specific representation vectors, and accurately recommend candidate suppliers based on similarity distance. As a result, not only the fit and timeliness of supplier matching results are significantly improved, but also manual intervention is effectively reduced, and the intelligence of procurement decisions and risk management capabilities are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0016] Figure 1 A flowchart of an example of a supplier matching method based on multi-source heterogeneous data fusion according to an embodiment of the present application is shown; Figure 2 An operational flow chart illustrating an example of dynamic updating of a supplier profile according to an embodiment of the present application is shown; Figure 3 An operational flowchart of an example of generating a supplier-specific representation vector based on an updated supplier profile according to an embodiment of the present application is shown; Figure 4 A schematic diagram showing the comparison of different methods on Top-M hit rate simulation; Figure 5 A structural block diagram of an example of a supplier matching system based on multi-source heterogeneous data fusion according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0017] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0018] It should be noted that in recent years, with the continuous expansion of construction project scale and the increasing complexity of procurement management, supplier matching has become a core component of mathematical analysis in bidding and tendering. Traditional construction supplier matching methods rely primarily on manual experience and static scoring. A common practice is for procurement personnel to manually screen or score suppliers based on sub-items such as supplier qualifications, historical cooperation records, quotations, and contract performance. While this method has a certain degree of operability, it is significantly subjective and fails to fully reflect the supplier's true capabilities and risk profile.

[0019] To improve matching efficiency and objectivity, some companies have introduced automated scoring systems based on databases and rule engines. These use pre-set rules to structure a variety of supplier data, including company size, financial status, contract fulfillment capabilities, and price competitiveness. At the same time, some research has proposed using multi-metric decision-making methods such as AHP (Analytic Hierarchy Process) and TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) for supplier evaluation and selection. These systems standardize and weight multiple metrics, partially enhancing the scientific nature of decision-making.

[0020] Furthermore, with the widespread adoption of information technology and big data, machine learning and data mining methods are gradually being incorporated into supplier matching. Related research attempts to build risk assessment and recommendation models based on historical procurement data, public opinion information, and third-party credit data. For example, these models employ decision trees, support vector machines, and cluster analysis to classify and predict suppliers, aiming to reduce labor costs and improve matching accuracy.

[0021] Despite continuous technological advancements, the following technical flaws are still prevalent in practical applications within the construction industry: First, supplier information suffers from significant data silos, making it difficult to effectively integrate data across projects, departments, and platforms. This results in incomplete supplier profiles and makes it difficult to dynamically reflect changes in suppliers' actual capabilities. Second, most existing models employ static scoring or single rules, making it difficult to adapt to dynamic matching requirements across multiple dimensions and scenarios, such as project type, project phase, and risk appetite. Third, current data-driven approaches often overlook the complex relationship and contextual adaptability between procurement needs and supplier capabilities, resulting in insufficient generalization and interpretability, which can easily lead to a disconnect between matching results and actual business practices. Finally, some AI-based models suffer from weak cold-start and online self-learning capabilities, making it difficult to adapt to new suppliers, new business models, or market changes in a timely manner.

[0022] Therefore, how to achieve the integration of multi-source heterogeneous data, dynamically optimize supplier portraits, and accurately adapt to procurement needs in multiple types of engineering scenarios remains a difficult problem that urgently needs to be broken through in current construction engineering supplier matching technology.

[0023] It should be understood that the purpose of the above description of the current related art is only to facilitate the public to better understand the inventive spirit and motivation of this application, and is not to be construed as limiting this application. In addition, the technical solutions described in the above-mentioned current related art are not prior art and may also be undisclosed technical solutions, such as solutions under research or in the laboratory stage.

[0024] Figure 1 A flowchart of an example of a supplier matching method based on multi-source heterogeneous data fusion according to an embodiment of the present application is shown, which constructs a highly timely intelligent matching technology system for construction project suppliers.

[0025] Regarding the executor of the method of the embodiment of the present application, it can be any controller or processor with computing or processing capabilities, which improves the intelligence and precision of supplier matching, realizes the dynamic response and differentiated adaptation of procurement decisions to multi-dimensional data, and promotes efficient collaboration and risk prevention in construction project supply chain management.

[0026] In some examples, it can be integrated into an electronic device or terminal through software, hardware, or a combination of software and hardware, and the type of terminal or electronic device can be diverse, such as a mobile phone, tablet computer, or desktop computer, etc.

[0027] like Figure 1 As shown, in step S110, feature analysis is performed on the engineering procurement requirement text according to a preset procurement feature dimension set to generate a corresponding procurement requirement vector.

[0028] It should be noted that the procurement requirement texts for engineering projects often come from a wide range of sources, including design task orders, bidding documents, user requirement specifications, etc., and their expressions are diverse and their emphases vary. Therefore, the original requirement texts need to be mapped into structured feature expressions.

[0029] Here, a pre-defined set of procurement characteristic dimensions provides a standardized handle for demand analysis, covering the core elements of procurement engineering. For example, these dimensions include any of the following: quality focus, timeliness of contract fulfillment, risk tolerance, and price sensitivity.

[0030] Quality focus reflects the purchaser's emphasis on product or service quality, such as whether it emphasizes high-standard construction and quality control at key milestones. Timeliness measures the sensitivity of procurement tasks to construction or delivery deadlines, such as whether a compressed construction period or phased delivery is required. Risk tolerance reflects the purchaser's tolerance for potential risks related to supplier creditworthiness, compliance, and litigation history. Price sensitivity reflects the extent of project budget constraints and the degree to which price is prioritized.

[0031] In some implementations, deep semantic parsing technology (such as the BERT language model) can be used to extract semantic elements from text and map them to a predefined set of procurement feature dimensions, forming a multidimensional procurement demand vector. Each dimension value represents the weight of that factor in the procurement text. For example, if a project emphasizes quality and timeliness but is insensitive to price, the dimension values ​​for "quality focus" and "performance timeliness" will be significantly higher than "price sensitivity." This achieves a standardized, structured, and quantitative expression of procurement requirements, generating fine-grained and hierarchical feature expressions for procurement requirements of different project types or with different priorities.

[0032] It should be understood that the above description of procurement feature dimensions is only used as an example to help the public more clearly understand the technical concepts and motivations of this application, and it should not be regarded as limiting the scope of implementation of this application.

[0033] Furthermore, procurement dimension sets can be customized to the project type, meaning different project types can have different sets of procurement dimension sets. For example, for rail transit projects, the procurement dimension sets might include contract fulfillment timeliness, compliance risk control, and safety standard compliance, while for residential construction projects, the dimensions might include quality focus, price sensitivity, and post-maintenance service capabilities.

[0034] In step S120, the real-time monitoring value of the supplier portrait indicator is determined based on multiple data sources, and when the deviation between the real-time monitoring value of the supplier portrait indicator and the corresponding indicator predicted value exceeds a preset amplitude threshold, the supplier portrait is updated by adjusting the portrait weight of the supplier portrait indicator.

[0035] It's important to note that supplier profiles, as a core input for supply-demand matching, have a significant impact on the reliability of recommended results due to their accuracy and timeliness. In engineering projects, a supplier's performance status can fluctuate significantly over time and within specific project circumstances. Traditional approaches rely solely on historical contract performance, qualification levels, and selected financial statements, often with lagging information updates and failing to reflect the supplier's latest dynamics.

[0036] In this embodiment, through multi-source data fusion and prediction deviation mechanism, the supplier portrait indicators are dynamically updated and the weights are adaptively adjusted, so that the portrait can timely and truly reflect the current and trend status of the supplier.

[0037] Here, the types of data sources can also be diverse, so as to obtain the real-time portrait indicator values ​​of suppliers (monitor real-time values) from numerous channels, such as the supplier's internal management system (such as the enterprise contract fulfillment system, quality supervision platform and project progress monitoring system, etc.), external data sources (such as third-party credit rating agency data, judicial litigation disclosure platform, public opinion monitoring tools, etc.) and financial data sources (supplier's financial statements, debt ratio, accounts receivable turnover cycle, etc.).

[0038] In addition, supplier profile indicators include any one of the following: fulfillment delay rate, quality pass rate, financial health index and judicial public opinion compliance.

[0039] The performance delay rate reflects the proportion of tasks that are not completed on time within the specified delivery time. The quality pass rate is calculated by counting the proportion of engineering nodes that have passed acceptance in the total number of acceptances. The financial health index is constructed by combining liquidity indicators and debt levels. The compliance with judicial public opinion can be comprehensively evaluated by combining administrative penalties, court judgments, negative online public opinion, etc.

[0040] It should be understood that the description of the above supplier portrait indicators is only used as an example to help the public more clearly understand the technical concepts and motivations of this application, and it should not be regarded as a limitation on the scope of implementation of this application.

[0041] After obtaining these real-time values, they are compared with the indicator forecasts stored in the supplier profile. These forecasts are calculated from the supplier profile's historical indicator data, such as those derived from a time series model using historical indicator data. When the deviation between the real-time and forecast values ​​exceeds a preset threshold (e.g., 15%), a weight adjustment mechanism is automatically triggered. For example, if a supplier's fulfillment delay rate suddenly increases significantly, the weight of that indicator in the overall profile will increase, shifting the overall profile toward "high risk" and enabling additional profile indicator warnings; the reverse is also true.

[0042] As a result, a continuous monitoring and dynamic feedback mechanism for the supplier's behavioral status is realized, and the changes in the supplier's performance capabilities are monitored in real time, which significantly enhances the consistency between the portrait and the actual status and avoids the risk misjudgment caused by the static and lagging portrait.

[0043] In step S130, a feature mapping relationship is selected according to the procurement project type, and the Top-K feature subsets with the highest correlation with the procurement demand vector are extracted from the updated supplier portrait, and weighted to generate a supplier-specific representation vector.

[0044] It should be noted that the indicators of supplier portraits are rich and diverse, but not every indicator is required to be assessed in procurement and bidding projects, and different types of engineering projects (such as municipal infrastructure, residential construction, industrial plants, etc.) have different emphases on supplier capability requirements. Therefore, by introducing a procurement type-driven feature mapping mechanism, key portrait features of the procurement project type are selected from the supplier portrait on demand for matching, thereby constructing an exclusive supplier representation vector.

[0045] For example, a predefined feature mapping relationship template is invoked based on the procurement project type (which can be derived from structured project type tags or parsed from procurement text). This template can be constructed based on historical project experience and expert knowledge, analyzing the sensitivity and weighting priorities of different project types to various supplier profile features. After the feature mapping relationship is determined, the top-K feature subset with the highest correlation (such as Pearson correlation coefficient or dot product similarity) with the current procurement requirement vector is selected from the supplier profile to ensure significant numerical coupling between the extracted features and the procurement requirements. Furthermore, the weight of each feature in the supplier-specific representation vector can be adjusted based on the corresponding dimension in the procurement requirement vector to achieve multi-dimensional alignment and semantic mapping.

[0046] In this embodiment, compared with the fixed full-feature matching method, dynamic mapping with the semantic content of the procurement task can effectively eliminate the interference of irrelevant or redundant features, and significantly improve the procurement task adaptability and focus of the supplier capability expression, thereby realizing customized supplier capability modeling.

[0047] In step S140 , the similarity distance between the supplier-specific representation vector and the procurement requirement vector is calculated to screen candidate suppliers.

[0048] Here, the similarity distance can be measured in a variety of ways, such as cosine similarity, Euclidean distance, Mahalanobis distance, etc. By calculating the similarity score between the procurement demand vector and each supplier's unique representation vector, personalized sorting of supplier matching results is achieved.

[0049] Through the embodiments of the present application, from customized procurement demand analysis of engineering types to exclusive coupling feature construction of dynamic supplier portraits, matching and sorting are performed based on vector similarity calculation results, which can more accurately identify the high degree of fit between suppliers and project needs, and can also keenly capture and respond to potential risks and demand changes, significantly improving the scientific nature and safety of decision-making in construction project supply chain management.

[0050] Regarding the implementation details of step S110, in some examples of the present application, the project procurement requirement text is extracted from multiple tender requirement documents published by the purchaser. Each tender requirement document indicates the same procurement project type and has a unique tender release time. Specifically, for each procurement feature dimension in a preset set of procurement feature dimensions, the characteristic time series value corresponding to that procurement feature dimension is extracted from the project procurement requirement text based on consecutive tender release times. Furthermore, a procurement requirement vector is generated based on the characteristic time series values ​​corresponding to each procurement feature dimension.

[0051] Here, by performing structured semantic analysis on all previous bidding texts of the same procurement project type, we extract the time series feature values ​​corresponding to the preset feature dimensions, and then combine these time series values ​​to generate a more trend-oriented dynamic procurement demand vector, which can reflect the purchaser's procurement preferences or changes in focus for the project.

[0052] It should be noted that construction engineering projects are usually characterized by long cycles, multiple stages, and repetitive types. The same type of procurement project often issues bidding documents multiple times at different time points. Although each document revolves around the same project goals, it may present differentiated demand expressions due to changes in the external environment, policy orientation, or project stages.

[0053] Therefore, by adopting the method of dynamic time series to generate procurement demand vectors and modeling based on the characteristic time series values ​​extracted from continuous bidding documents, it is possible to more objectively and systematically reflect the purchaser's long-term focus and emphasis on this type of project, realize the modeling of the purchaser's long-term behavioral preferences, and at the same time improve the model's responsiveness to the evolution of procurement intentions.

[0054] For example, in the procurement project category of "Urban Rail Transit Tunnel Construction," a company has issued six tender requirements documents over the past three years, spanning Q1 2021 to Q2 2024. Systematic processing revealed the following trends: Quality focus consistently remained above 0.85, indicating the company's adherence to high construction standards; contract fulfillment timelines significantly improved in 2023 due to the overall shortened subway line cycle; price sensitivity briefly increased initially before declining; and risk tolerance continued to decline, reflecting tightening regulations. This allows us to dynamically represent the weighting of different procurement feature dimensions, resulting in a richer and more precise semantic representation of the procurement demand vector.

[0055] It should be noted that in the actual procurement scenarios of construction projects, the level of detail of the content of the bidding requirements text varies greatly, resulting in the inability to fully extract the information required for the preset procurement feature dimension set (such as quality attention, fulfillment time, risk tolerance, and price sensitivity) from certain texts during structured analysis.

[0056] Given this, some examples of the present application's embodiments test whether the project procurement requirement document covers each procurement feature dimension in the procurement feature dimension set. If so, a procurement requirement vector can be directly generated; if not, this indicates a missing dimension. In this case, a historical project procurement data set matching the procurement project type can be obtained, and the feature averages corresponding to the missing procurement feature dimensions in the historical project procurement data set can be extracted. The missing procurement feature dimensions are then supplemented based on the feature averages. Furthermore, a procurement requirement vector is generated based on the supplemented procurement feature values.

[0057] In this embodiment of the application, the current procurement requirement text is checked for dimensional coverage. If there are missing dimension items, it is impossible to directly construct a complete procurement requirement vector. To this end, a historical engineering procurement data set that matches the current project type can be called up, and the feature averages of these historical data on the missing dimensions can be extracted as approximate expressions of the corresponding dimensions in the current requirement vector. The dimension values ​​obtained from the test are merged with the missing dimension values ​​to construct a complete procurement requirement vector.

[0058] For example, if "risk tolerance" is missing from the current document, the system extracts the eigenvalues ​​for that dimension from historical similar projects, such as [0.42, 0.48, 0.44, 0.46], and takes the average of 0.45 as the missing estimate for "risk tolerance." This ensures that even with incomplete information in the engineering procurement requirements document, a stable vector representation is maintained, preventing the entire supplier matching process from being impacted by missing dimensional values.

[0059] Regarding the details of screening candidate suppliers by calculating the similarity distance, in some examples of the embodiments of the present application, the supplier-specific representation vector and the procurement demand vector can be dimensionally mapped in the differentiated feature subset space. Here, the procurement demand vector and the vendor-specific representation vector Both are based on differentiated feature subsets Differentiated feature subsets constructed in the same dimensional space ( ) is selected based on the feature relevance scoring results of the procurement requirements and corresponds to the supplier portrait dimension. This ensures that when calculating the similarity, the two vectors are aligned in the same feature dimension space without the risk of dimensional misalignment.

[0060] Based on the normalized weight of each feature in the differentiated feature subset, the weighted Euclidean distance and weighted cosine similarity between the supplier-specific representation vector and the procurement demand vector are calculated respectively.

[0061] Here, each feature is weighted using normalized weights to calculate weighted Euclidean distance and weighted cosine similarity. Weighted Euclidean distance measures the absolute deviation between supplier capabilities and procurement requirements, emphasizing the degree of fit between the two along the same feature dimension. Furthermore, weighted cosine similarity measures the directional similarity between supplier capabilities and procurement requirements in feature space, making it particularly suitable for structural fit between supplier capabilities and requirements.

[0062] In the calculation of weighted Euclidean distance, the portrait vector of each supplier is and the purchase demand vector , according to the normalized weight For each feature dimension Compute the squares of the weighted differences and sum them: , formula (1) Where, Characterized by The normalized weight of For suppliers In Features The standardized value on The purchasing demand vector is characterized by The demand value on .

[0063] In the calculation of weighted cosine similarity, the supplier portrait vector and the purchase demand vector , calculate the weighted cosine similarity based on the normalized weights of the features: , formula (2) Therefore, the absolute fit between suppliers and demand is measured by weighted Euclidean distance, especially when certain features (such as cost and delivery time) need to be strictly controlled; weighted cosine similarity focuses more on the matching of feature structures and can identify the directional fit between suppliers and demand. It is especially suitable for scenarios with diverse demands and emphasis on long-term cooperative relationships.

[0064] The weighted Euclidean distances of all suppliers are normalized to obtain the distance matching score, which is then weightedly fused with the weighted cosine similarity according to the preset fusion weight to generate a comprehensive supplier similarity score.

[0065] In order to ensure fair comparison between different measurement methods, the weighted Euclidean distance and weighted cosine similarity need to be normalized so that their value ranges are consistent and easy to fusion calculation. The normalized weighted Euclidean distance reflects the "absolute difference" between suppliers and demand; the normalized weighted cosine similarity reflects the "directional fit" between suppliers and demand. After that, the two are weighted and fused according to the preset fusion weight (for example, 0.5, or dynamically adjusted according to demand) to obtain a comprehensive similarity score. Through normalization and weighted fusion, the dimensional differences between different similarity measurement methods are eliminated.

[0066] Finally, all suppliers are sorted according to the comprehensive similarity scores of suppliers and preset screening rules to screen candidate suppliers.

[0067] Here, all suppliers are ranked based on their comprehensive similarity score. A higher score indicates a supplier's greater suitability for the current procurement requirements. Therefore, you can set a minimum match score threshold or a Top-N filter based on your needs to ensure that recommended suppliers meet the purchaser's core requirements.

[0068] It should be noted that matching suppliers with procurement requirements does not rely solely on the fit of a single feature, but rather involves a comprehensive assessment of multiple dimensions. For example, the differences between a supplier's capabilities and procurement requirements (such as fulfillment timeliness and quality control) are crucial, so using weighted Euclidean distance can quantify these absolute numerical differences. At the same time, procurement requirements often focus on structural fit, such as whether the supplier's capabilities and the required characteristics match in the same direction, which requires weighted cosine similarity to measure. Therefore, by combining weighted Euclidean distance and weighted cosine similarity, we can comprehensively consider both absolute fit and directional fit, ensuring that the recommendation system can more comprehensively assess the fit between suppliers and procurement requirements.

[0069] It should be noted that in the analysis of construction project bidding and procurement, the actual performance behavior of suppliers is highly dynamic and time-sensitive. Key portrait indicators such as their performance quality, delivery capability, and compliance level may fluctuate over time due to environmental changes, resource pressure, policy supervision and other factors. Static portrait models often cannot reflect these changes in real time, which may lead to matching deviations and risk exposure.

[0070] In view of this, Figure 2 An operational flowchart of an example of dynamic updating of supplier portraits according to an embodiment of the present application is shown, and a supplier portrait updating method based on historical prediction-real-time observation comparison and weight adaptive control mechanism is proposed to improve the dynamic perception ability and expression reliability of the portrait.

[0071] like Figure 2As shown, in step S210, for each supplier portrait indicator, the historical time series value of the supplier portrait indicator is processed based on the exponentially weighted moving average method to obtain the corresponding indicator prediction value.

[0072] Supplier profile indicators, as important quantitative data for measuring a supplier's performance (such as fulfillment delay rate, quality compliance rate, financial health index, and compliance with legal and public opinion), typically exhibit certain time series characteristics. To determine whether current supplier behavior is "deviant," a reasonable historical trend baseline must first be established. Specifically, the historical time series values ​​of each profile indicator are processed using the EWMA (Exponentially Weighted Moving Average) model to obtain its predicted value at the current moment, which serves as a reference baseline for "normal status."

[0073] Specifically, for each supplier profile indicator, the system maintains a time series consisting of observation data at consecutive time points, denoted as , Represents the historical value of the indicator at time point t.

[0074] Use the EWMA algorithm to generate forecast values: , formula (3) Where, represents the predicted value of the indicator at time point t, Represents the real historical observation value at time point t-1; is the smoothing factor, the larger the value, the more sensitive it is to recent data; It is the predicted value at time point t-2. The initial value can be the historical average or the first observation value.

[0075] In order to adapt to the dynamics of different indicators, the system allows different For example, for an indicator with high volatility and significant short-term changes such as "delay rate of fulfillment", a larger value can be selected. (e.g., 0.6~0.8); for indicators that change slowly, such as the "Financial Health Index", a smaller value may be selected. (For example, 0.2-0.4). As a result, the generated forecast value has the characteristics of both trend smoothness and recent sensitivity, effectively avoiding misjudgments caused by short-term fluctuations and enhancing the system's ability to grasp the indicator's evolution trend.

[0076] In step S220, original observation values ​​of supplier portrait indicators are obtained from multiple data sources, and real-time monitoring values ​​of supplier portrait indicators are obtained through weighted fusion.

[0077] To truly understand the performance of a supplier's current profile, it's necessary to collect raw observations from multiple business platforms and data sources. These sources include internal management systems (such as project management platforms and quality inspection systems) and external platforms (such as third-party credit ratings, public opinion systems, and court announcement platforms). Due to the heterogeneity and varying reliability of data sources, a weighted fusion mechanism must be introduced to integrate observations and generate more stable and reliable real-time monitoring values.

[0078] For example, suppose a certain portrait indicator is The observations collected from different data sources are , set weights for each observation , satisfying the normalization condition: , then the real-time monitoring value of the portrait indicator The calculation formula is: , formula (4) Where, Indicates the The observation values ​​of the data source, Indicates the The fusion weight of each data source can be set based on factors such as data credibility and update frequency.

[0079] Therefore, by integrating multi-source observation data, the accuracy and real-time performance of portrait indicator monitoring can be significantly improved, avoiding the delays or false alarms that may be caused by relying on a single source, and enhancing the system's ability to perceive the actual status of suppliers.

[0080] In step S230, when it is detected that the deviation between the real-time monitoring index value of the supplier portrait index and the index prediction value exceeds the preset amplitude threshold, the weight of the corresponding supplier portrait index is adjusted according to the deviation amplitude.

[0081] On the other hand, when the preset amplitude threshold is not exceeded, there is no need to adjust the weight of the supplier profile indicator.

[0082] Here, if there is a significant deviation between the real-time monitoring value and the historical forecast value, it means that there is an abnormal change in the performance of the portrait indicator in the current period. Its weight should be adjusted to strengthen its influence on the overall portrait, thereby increasing the sensitivity of the portrait to actual risks; conversely, if the deviation is acceptable, no adjustment will be made.

[0083] Specifically, suppose the predicted value of an indicator is , the real-time monitoring value is , calculate its relative deviation : , formula (5) System-set default deviation threshold , the judgment conditions are: like , then the indicator is considered to be significantly abnormal and its weight should be adjusted; if , it is regarded as normal fluctuation and no adjustment is made.

[0084] For indicators that need to adjust their weights, their new weights It is calculated as follows: , formula (6) Where, is the old weight of the indicator; It is an adjustment coefficient that controls the sensitivity of deviation amplification (a setting of 1 indicates a linear response).

[0085] In addition, to prevent the weight from expanding too quickly, the system can set a maximum weight limit To crop: , formula (7) Unlike the traditional fixed-weight portrait model, the embodiment of the present application can dynamically respond to abnormal fluctuations in supplier portrait indicators, and by enhancing their weights in the portrait, the portrait results can quickly reflect current risk points or potential performance issues, thereby enhancing the adaptability and risk warning capabilities of the portrait.

[0086] In step S240, the portrait weights of all supplier portrait indicators are normalized to update the supplier portrait.

[0087] Here, after the profile weight adjustment in step S230, the sum of the weights of each profile indicator may change, no longer meeting the weighted consistency requirement. To maintain profile comparability and structural integrity, all indicator weights must be normalized and the supplier profile updated accordingly. This ensures that the sum of the dimension weights of the supplier profile is 1, maintaining structural consistency and avoiding imbalances caused by local weight adjustments.

[0088] Figure 3 An operational flowchart of an example of generating a supplier-specific representation vector based on an updated supplier portrait according to an embodiment of the present application is shown.

[0089] like Figure 3 As shown, in step S310, according to the procurement project type identification , load the pre-trained engineering type-feature mapping matrix ,in is the number of associated procurement feature dimensions, which is equal to the total number of dimensions in the procurement feature dimension set. The total number of associated supplier profile metrics calibrated for pre-training.

[0090] It should be noted that the rows of the feature map matrix (i.e. Dimensions) are consistent with the procurement feature dimension set. This means that dimension misalignment or information loss will not occur during feature extraction and mapping, which helps ensure the traceability and explainability of procurement requirements. Furthermore, improving feature relevance and business matching ensures a tight coupling between the feature mapping matrix and procurement requirements. The selected differentiated feature subset (i.e., the Top-K features) is guaranteed to be of actual interest to procurement, improving the accuracy of the results.

[0091] In some embodiments, the system stores a set of pre-trained project type-feature mapping matrices, which record the feature mapping matrices corresponding to different project types. Based on the type of project procurement requirement (e.g., construction, equipment procurement, road construction, etc.), the system loads a matching pre-trained model. Each pre-trained model contains a feature mapping matrix specific to that project type, which is used to map the various features of the supplier profile into a high-dimensional space. This ensures that the model training and feature mapping are tailored to the requirements of the specific project type, avoiding mismatches caused by generic feature mappings.

[0092] For example, assuming that characteristics related to procurement needs (such as quality, timeliness of fulfillment, risk, etc.), and Supplier characteristic dimensions, the system will use engineering type The corresponding feature map matrix Feature mapping is performed to map the characteristics of each supplier into a feature space related to the project type to ensure that the relevance to procurement requirements is maximized.

[0093] In step S320, according to the purchase demand vector and engineering type-feature mapping matrix Compute feature relevance score vector.

[0094] , formula (8) Where, express The transposed vector of Represents the feature relevance score vector.

[0095] Specifically, Equation (8) reflects a linear feature weighted projection. Represents the attention and quantitative requirements of this procurement on different features (such as price, delivery time, quality, etc.), and the project type-feature mapping matrix This describes the structured mapping relationship between each procurement feature and all supplier profile indicators. Multiplying the two is equivalent to weighting the correlation strength between each feature and all supplier profile characteristics under each project type by the importance of each procurement requirement feature, thus achieving a "weighted correlation assessment" of each supplier profile indicator.

[0096] That is, the result of matrix multiplication is Each component in the metric represents the contribution or relevance of each profile feature to satisfying the current project's requirements, given the current procurement demand's focus weight. This allows the system to automatically quantify which supplier capabilities are most critical in the current demand context, enabling targeted supplier screening and recommendation. This linear mapping approach quantifies the correlation between current project requirements and each supplier's profile features while maintaining business interpretability.

[0097] In step S330, select The largest value The index of the supplier portrait indicators constitutes a subset of differentiated features .

[0098] Here, after obtaining the matching scores between suppliers and procurement requirements, the system will sort these scores and select the one with the highest score. supplier profile indicators to highlight the most relevant profile indicators while avoiding interference from low-correlation or noise dimensions.

[0099] Specifically, the system scores the vector according to the feature relevance , select the front An index collection of supplier profile indicators , indicating the one that best matches the current demand In addition, you can also adjust The value can flexibly control the number of candidate supplier profile indicators to meet the actual needs of different business scenarios or system resources.

[0100] In step S340 , a supplier-specific representation vector is extracted from the updated supplier profile based on the differentiated feature subset.

[0101] In some embodiments, after selecting the optimal After generating the supplier portrait indicators, the system further extracts feature vectors of the values ​​under these indicators to reflect the supplier's adaptability under the current procurement needs.

[0102] The system updates the vendor-specific representation vector using the following formula: , formula (9) , formula (10) Where, Indicates supplier The vendor-specific representation vector of For suppliers Indicators in supplier profile The standardized index value of Profiling indicators for suppliers The normalized weight of The influence of a feature within a subset of differentiated features; is the adjustment coefficient, which is used to adjust the concentration and dispersion of weight distribution; is a normalized ergodic variable, representing the index of any feature in the differentiated feature subset; Represents the first in the differential feature subset The relevance score of a feature.

[0103] This embodiment of the application introduces a feature mapping matrix based on procurement project type. By operating on the procurement demand vector, it identifies the most relevant subset of profile indicators for the current task and performs a weighted aggregation of these indicators within each supplier. Furthermore, the weights are adaptively adjusted based on the relevance score for the actual procurement demand, allowing the profile-specific vector to focus more closely on the capability dimensions of interest to the task, helping to improve the accuracy of the matching ranking results.

[0104] Regarding the training of the engineering type-feature mapping matrix set, in some examples of the embodiments of the present application, it can be achieved through the following operations.

[0105] First, from the project data of multiple historical engineering types, we collect the procurement demand vector set, supplier portrait indicator set, and supplier matching results of historical projects to initialize the mapping matrix element parameters shared by all engineering types. .

[0106] It should be understood that the effectiveness of the project type-feature mapping matrix is ​​highly dependent on the comprehensiveness and accuracy of three types of historical data: procurement requirements, supplier profiles, and actual matching results. For example, from historical project data for multiple project types (such as roads, bridges, and residential buildings), we collect and organize, in batches, each project's procurement requirement vector set (e.g., required materials, technical parameters), supplier profile indicator sets (e.g., performance capability, cost, reputation), and the actual matching results between the project and different suppliers (e.g., winning bids, ratings, contract performance, etc.).

[0107] By processing these data in a structured manner, procurement requirements and supplier capabilities can be uniformly mapped into a standardized feature space, providing data support for multi-type matrix training. On this basis, a global meta-parameter is initialized through the historical sample averaging method. , as the initial parameters of the mapping matrix of all engineering types. Therefore, the unified initialization of meta-parameters can effectively improve the efficiency of knowledge transfer and parameter sharing between different engineering types.

[0108] Then, for each project type, the procurement demand vector, supplier profile index and historical project matching results of the project type are used to calibrate the meta-parameters. Perform a preset number of gradient updates to generate the initial engineering type-feature mapping matrix : , formula (11) Where, is the learning rate, For project types The loss function is defined by combining procurement requirements, supplier profiles and historical matching results.

[0109] It should be noted that the requirements and supplier capabilities of each project type have different focuses, and global meta-parameters need to be Only by performing type-customized fine-tuning can we obtain an exclusive mapping matrix that reflects the characteristics of this project type.

[0110] Specifically, for each specific project type , the procurement demand vector, supplier portrait indicators and historical matching result data related to this type are selected as the training set.

[0111] Construct loss functions for each engineering type-feature mapping matrix model The project type-feature mapping matrix model can employ a linear mapping model or a neural network, and is used to optimize the feature mapping matrix for the corresponding project type. This allows the actual matching effect between procurement requirements and supplier profiles to be predicted as accurately as possible after matrix mapping. Therefore, this loss function comprehensively reflects the deviation between the matching effect between procurement requirements and supplier profiles and historical observations. It can employ mean squared error, cross entropy, or a weighted loss function customized for each project type. For example, the mean squared error loss function measures the mean squared error between the model-predicted supplier matching scores and the actual matching results for all historical samples.

[0112] by Based on this, the gradient descent method is used to iterate the loss function once or multiple times, and the parameters are updated to minimize the matching deviation to obtain the exclusive feature mapping matrix of this engineering type. .

[0113] This enables the feature mapping matrices of different engineering types to accurately express the correlation between procurement characteristics and supplier profiles for that type, improving the accuracy and pertinence of intelligent supplier matching. Furthermore, it supports parameter sharing and knowledge transfer between different types, improving the robustness and generalization of new type matrix training in small sample scenarios.

[0114] Then, by jointly optimizing the sum of fine-tuned losses under all engineering types, the following objective function is minimized to obtain the meta-parameters with the best generalization ability: : , formula (12) Where, Indicates the total number of historical project types, Represents a pair parameter In engineering type Loss function The gradient of Adjustments to reduce engineering types matching loss.

[0115] Here, in order to further improve the generalization ability of the model across different engineering types, the multi-task learning idea is adopted to jointly minimize the sum of the fine-tuned losses of all engineering types to obtain the globally optimal meta-parameters. Therefore, meta-learning and automatic differentiation are used to Multi-task training makes it the optimal starting point for fine-tuning matrices specific to each engineering type.

[0116] When updated project data is detected, the meta parameters are fine-tuned using the updated project data. , and update the engineering type-feature mapping matrix set.

[0117] Here, an online adaptive update mechanism is adopted. When new project data (such as new procurement requirements, supplier profiles, and the latest matching / performance performance) arrives in real time, these updated project data are used as increments, and the same gradient update or fine-tuning strategy as historical data is adopted to adjust the global meta-parameters. Or a dedicated matrix for the corresponding project type Perform local optimization. As a result, the mapping matrix can continuously track and reflect market dynamics, reducing the risk of adaptability degradation or failure caused by long-term reliance on static models.

[0118] Through the embodiments of the present application, based on historical procurement needs, supplier profiles, and actual matching results, a unique feature mapping matrix that closely corresponds to historical performance is trained for each project type. In addition, efficient knowledge transfer between different types of models is achieved through multi-task optimization of global meta-parameters, reducing the difficulty of cold starting for new types or data-scarce types. In addition, through multi-task joint optimization and continuous online updates, the system can learn and adaptively reflect the actual business laws in various engineering scenarios.

[0119] In order to further verify the superiority of the method of the embodiment of the present application, the details of the experimental part of the method based on the engineering type-feature mapping matrix model of the embodiment of the present application will be expanded below.

[0120] Specifically, to validate the effectiveness and superiority of the proposed "supplier matching method based on a project type-feature mapping matrix" in real-world engineering procurement scenarios, this paper conducted experiments using procurement datasets from multiple real-world construction projects. The experimental data encompasses procurement demand characteristics, supplier profile indicators, and historical matching scores. The data covers multiple project types, including bridges, residential buildings, and roads, and encompasses varying project sizes, complexities, and supplier types.

[0121] In data preprocessing, all features were standardized and labeled as historical matching scores (or fulfillment performance scores) between procurement requirements and supplier profiles. The dataset was randomly divided into training and test sets to ensure the objectivity of the experimental results.

[0122] In terms of evaluation indicators, the following mainstream evaluation indicators are used to evaluate the matching effect of each method: 1) Mean Squared Error (MSE): reflects the error between the model's predicted score and the actual score.

[0123] 2) Mean Absolute Error (MAE): measures the average deviation between the predicted value and the true value.

[0124] 3) Top-M hit rate: In the test set procurement requirements, whether the actual preferred supplier is listed in the model's top-M recommendation results (measures ranking and recommendation accuracy).

[0125] 3) AUC: reflects the consistency and discrimination of model ranking.

[0126] To comprehensively compare the performance of the proposed method, two existing models were selected for comparative experiments. The first is a traditional static weighted model, which sets weights based solely on manual experience or fixed rules and performs linear weighted matching based on each feature, regardless of project type. The second is a standard multi-index linear regression model, which uses a standard linear regression model to directly concatenate procurement requirements and supplier profile features as input and outputs a matching score without using a mapping matrix structure.

[0127] The two comparative models above are compared with the matching method based on the project type-feature mapping matrix model proposed in this paper. Specifically, the project type-feature mapping matrix model introduces project type differentiation and uses the feature mapping matrix training and fine-tuning mechanism proposed in this paper to correlate and project procurement requirements and supplier profiles, and intelligently match them.

[0128] Table 1 shows the main evaluation indicators of different methods on the test set

[0129] As shown in the experimental results in Table 1, our approach outperforms traditional methods and standard linear regression models across all evaluation metrics, particularly achieving significant improvements in key metrics such as Top-5 hit rate and Area Under Current Arbitrage (AUC). Therefore, by introducing the project type-feature mapping matrix mechanism, we can effectively model the complex correspondence between procurement requirements and supplier profiles for different project types, enabling personalized, intelligent, and accurate supplier matching.

[0130] Figure 4 The following is a schematic diagram showing the comparison of different methods on Top-M hit rate. Figure 4 A simulation compares the Top-M hit rates of a project type-feature mapping matrix (method proposed in this paper), a multi-index linear regression model, and a traditional static weighted model on the same test set. The horizontal axis represents the length of the recommended list (M), ranging from 1 to 10 (i.e., the Top-M hit rate). The vertical axis represents the corresponding Top-M hit rate (%), which indicates the proportion of historically preferred suppliers appearing in the model's top M recommendations.

[0131] like Figure 4 The results shown are for the static weighted model (dotted line): each feature is linearly weighted using a fixed empirical weight, and the hit rate increases slowly as M increases; Multi-index linear regression (square line): This method directly combines procurement requirements with supplier profile features and uses linear regression to predict matching scores. Its Top-M hit rate is higher than that of static methods. The method in this paper (triangle line): an intelligent matching scheme based on the project type-feature mapping matrix, significantly outperforms the comparison method under all M values, especially when M=5, with a hit rate of over 82%. This verifies the method's ability to efficiently model the complex correspondence between procurement requirements and supplier profiles under different project types.

[0132] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of combined actions, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0133] Figure 5 A structural block diagram of an example of a supplier matching system based on multi-source heterogeneous data fusion according to an embodiment of the present application is shown.

[0134] like Figure 5 As shown, the supplier matching system 500 based on multi-source heterogeneous data fusion includes a procurement demand analysis unit 510, a supplier portrait update unit 520, a Top-K vector generation unit 530 and a supplier screening unit 540.

[0135] The procurement requirement analysis unit 510 is used to perform feature analysis on the engineering procurement requirement text according to a preset set of procurement feature dimensions to generate a corresponding procurement requirement vector; the procurement feature dimensions include any one of the following: quality concern, fulfillment timeliness, risk tolerance and price sensitivity.

[0136] The supplier portrait update unit 520 is used to determine the real-time monitoring value of the supplier portrait indicator based on multiple data sources, and when the deviation between the real-time monitoring value of the supplier portrait indicator and the corresponding indicator prediction value exceeds a preset amplitude threshold, the supplier portrait is updated by adjusting the portrait weight of the supplier portrait indicator; the indicator prediction value is a real-time prediction value calculated based on the historical indicator data of the supplier portrait indicator, and the supplier portrait indicator includes any one of the following: fulfillment delay rate, quality pass rate, financial health index and judicial public opinion compliance.

[0137] The Top-K vector generation unit 530 is used to select feature mapping relationships according to the procurement project type, extract the Top-K feature subsets with the highest correlation with the procurement demand vector from the updated supplier portrait, and weightedly generate a supplier-specific representation vector.

[0138] The supplier screening unit 540 is configured to calculate the similarity distance between the supplier-specific representation vector and the procurement requirement vector to screen candidate suppliers.

[0139] In some embodiments, an embodiment of the present application provides a non-volatile computer-readable storage medium, which stores one or more programs including execution instructions, and the execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute any of the steps of the above-mentioned supplier matching method based on multi-source heterogeneous data fusion in the present application.

[0140] In some embodiments, the embodiments of the present application also provide a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer is caused to perform any step of the above-mentioned supplier matching method based on multi-source heterogeneous data fusion.

[0141] In some embodiments, an embodiment of the present application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of a supplier matching method based on multi-source heterogeneous data fusion.

[0142] The above-mentioned product can execute the method provided in the embodiment of this application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of this application.

[0143] The electronic devices of the embodiments of the present application exist in various forms, including but not limited to: (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and their primary purpose is to provide voice and data communications. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0144] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers and have computing and processing capabilities, and generally also have mobile Internet access. These terminals include PDAs, MIDs, and UMPCs.

[0145] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players, handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0146] (4) Other onboard electronic devices with data interaction functions, such as onboard computer devices installed in vehicles.

[0147] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0148] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a general hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the relevant technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A supplier matching method based on multi-source heterogeneous data fusion, characterized in that: The method comprises: Perform feature parsing on the engineering procurement requirement text based on a preset set of procurement feature dimensions to generate a corresponding procurement requirement vector; procurement feature dimensions include any one of the following: quality focus, contract fulfillment timeliness, risk tolerance, and price sensitivity; Determine the real-time monitoring value of the supplier portrait indicator based on multiple data sources, and when the deviation between the real-time monitoring value of the supplier portrait indicator and the corresponding indicator forecast value exceeds a preset amplitude threshold, update the supplier portrait by adjusting the portrait weight of the supplier portrait indicator; the indicator forecast value is a real-time forecast value calculated based on the historical indicator data of the supplier portrait indicator, and the supplier portrait indicator includes any one of the following: fulfillment delay rate, quality pass rate, financial health index, and judicial public opinion compliance; Select feature mappings based on the procurement project type, extract the top-K feature subsets with the highest correlation with the procurement demand vector from the updated supplier profile, and weight them to generate a supplier-specific representation vector. The similarity distance between the supplier-specific representation vector and the procurement demand vector is calculated to screen candidate suppliers.

2. The method according to claim 1, characterized in that The engineering procurement requirement text is extracted from multiple bidding requirement documents issued by the purchaser, each of which is used to indicate the same type of procurement project and has a unique corresponding bidding release time; The feature parsing of the engineering procurement requirement text according to the preset procurement feature dimension set to generate a corresponding procurement requirement vector includes: For each procurement feature dimension in the preset procurement feature dimension set, extracting a feature time series value corresponding to the procurement feature dimension from the engineering procurement requirement text according to consecutive bidding release times; A procurement demand vector is generated according to the characteristic time series values ​​corresponding to each of the procurement characteristic dimensions.

3. The method according to claim 1, characterized in that The feature parsing of the engineering procurement requirement text according to the preset procurement feature dimension set to generate a corresponding procurement requirement vector includes: Detecting whether the engineering procurement requirement text covers each procurement feature dimension in the procurement feature dimension set; If there is no coverage, obtain a historical engineering procurement data group that matches the procurement engineering type, extract the feature average value of the missing procurement feature dimension corresponding to the historical engineering procurement data group, and complete the missing procurement feature dimension based on the feature average value; Generate a procurement demand vector based on the improved procurement feature values.

4. The method according to claim 1, wherein The method of determining the real-time monitoring value of the supplier portrait indicator based on multiple data sources and updating the supplier portrait by adjusting the portrait weight of the supplier portrait indicator when the deviation between the real-time monitoring value of the supplier portrait indicator and the corresponding indicator predicted value exceeds a preset amplitude threshold includes: For each supplier portrait indicator, the historical time series value of the supplier portrait indicator is processed based on the exponentially weighted moving average method to obtain the corresponding indicator forecast value; Obtain the original observation values ​​of supplier profile indicators from multiple data sources, and obtain the real-time monitoring values ​​of supplier profile indicators through weighted fusion; When it is detected that the deviation between the real-time monitoring index value of the supplier profile index and the index prediction value exceeds the preset threshold, the weight of the corresponding supplier profile index is adjusted according to the deviation; Normalize the portrait weights of all supplier portrait indicators to update the supplier portrait.

5. The method according to claim 4, characterized in that The feature mapping relationship is selected according to the procurement project type, and the top-K feature subsets with the highest correlation with the procurement demand vector are extracted from the updated supplier profile, and the supplier-specific representation vector is generated by weighting, including: Identify by procurement project type , load the pre-trained engineering type-feature mapping matrix ,in is the number of associated procurement feature dimensions, which is equal to the total number of dimensions in the procurement feature dimension set. The total number of associated supplier profile indicators calibrated for pre-training; According to the purchasing demand vector and engineering type-feature mapping matrix Calculate the feature relevance score vector: ， Where, express The transposed vector of represents the feature relevance score vector; Select The largest value The index of the supplier portrait indicators constitutes a subset of differentiated features ； Extract supplier-specific representation vectors from the updated supplier profile based on the differentiated feature subset: ， ， Where, Indicates supplier The vendor-specific representation vector of For suppliers Indicators in supplier profile The standardized index value of Profiling indicators for suppliers The normalized weight of The influence of a feature within a subset of differentiated features; is the adjustment coefficient, which is used to adjust the concentration and dispersion of weight distribution; is a normalized ergodic variable, representing the index of any feature in the differentiated feature subset; Represents the first in the differential feature subset The relevance score of a feature.

6. The method according to claim 5, characterized in that The training for the engineering type-feature map matrix set includes: From the project data of multiple historical engineering types, we collect procurement demand vector sets, supplier profile indicator sets, and supplier matching results of historical projects to initialize the mapping matrix element parameters shared by all engineering types. ; For each project type, the meta parameters are compared using the procurement demand vector, supplier profile index and historical project matching results of the project type. Perform a preset number of gradient updates to generate the initial engineering type-feature mapping matrix : ， Where, is the learning rate, For project types The loss function is defined by combining procurement requirements, supplier profiles and historical matching results; By jointly optimizing the sum of fine-tuned losses under all engineering types, the following objective function is minimized to obtain the meta-parameters with the best generalization ability: : ， Where, Indicates the total number of historical project types, Represents a pair parameter In engineering type Loss function The gradient of Adjustments to reduce engineering types Matching loss; In the event that updated project data is detected, the meta-parameters are fine-tuned using the updated project data. , and update the engineering type-feature mapping matrix set.

7. The method according to claim 5, characterized in that Calculating the similarity distance between the supplier-specific representation vector and the procurement requirement vector to screen candidate suppliers includes: Dimensionally mapping the supplier-specific representation vector and the procurement demand vector in the differentiated feature subset space; Based on the normalized weight of each feature in the differentiated feature subset, respectively calculating the weighted Euclidean distance and weighted cosine similarity between the supplier-specific representation vector and the procurement demand vector; Normalizing the weighted Euclidean distances of all suppliers to obtain a distance matching score, and performing weighted fusion with the weighted cosine similarity according to a preset fusion weight to generate a comprehensive supplier similarity score; According to the comprehensive similarity scores of the suppliers and the preset screening rules, all suppliers are sorted to screen candidate suppliers.

8. A supplier matching system based on multi-source heterogeneous data fusion, characterized by: The system comprises: A procurement demand analysis unit is used to perform feature analysis on the engineering procurement demand text based on a preset set of procurement feature dimensions to generate a corresponding procurement demand vector; the procurement feature dimensions include any one of the following: quality focus, timeliness of contract fulfillment, risk tolerance, and price sensitivity; A supplier portrait updating unit, configured to determine a real-time monitoring value of a supplier portrait indicator based on multiple data sources, and to update the supplier portrait by adjusting the portrait weight of the supplier portrait indicator when the deviation between the real-time monitoring value of the supplier portrait indicator and the corresponding indicator predicted value exceeds a preset amplitude threshold; the indicator predicted value is a real-time predicted value calculated based on historical indicator data of the supplier portrait indicator, and the supplier portrait indicator includes any one of the following: fulfillment delay rate, quality pass rate, financial health index, and judicial public opinion compliance; A Top-K vector generation unit is used to select a feature mapping relationship based on the procurement project type, extract the Top-K feature subset with the highest correlation with the procurement demand vector from the updated supplier profile, and generate a supplier-specific representation vector based on the weights; The supplier screening unit is used to calculate the similarity distance between the supplier-specific representation vector and the procurement demand vector to screen candidate suppliers.

Images