Supplier grading and classifying method based on multi-dimensional feature fusion

Abstract

The invention provides a supplier grading and classifying method based on multi-dimensional feature fusion, and relates to the technical field of supply chain management. The method comprises the steps of obtaining structured indexes and unstructured text data of suppliers; structural features are extracted through a multi-layer perceptron, and semantic features are extracted through Transform; a cross-modal attention mechanism is adopted, and deep fusion is carried out with the structured features as queries and the semantic features as key values to obtain interaction features; performing dynamic modeling on the historical time sequence data through bidirectional LSTM to generate time sequence comprehensive features; inputting a multi-task hierarchical classifier, combining ordinal regression and contrast learning loss optimization, and outputting categories and grades; according to the management priority, the weight is dynamically adjusted through attention bias, and grading flexible adaptation is achieved; and incremental learning fine tuning is carried out by using actual service data through closed-loop feedback. The method solves the problems that multi-source heterogeneous features are difficult to fuse, and feature space semantic distinguishing and grading standards lack dynamic adaptation, and is applied to intelligent evaluation of industrial internet suppliers.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, specifically to a supplier classification and grading method based on multi-dimensional feature fusion. Background Technology

[0002] Early supplier classification practices relied heavily on the industry experience of purchasing personnel or simple ranking based on single business indicators. While this approach was somewhat intuitive when dealing with small-scale data, its limitations became increasingly apparent when facing modern, large-scale, and heterogeneous supplier data, including its strong subjectivity, slow response, and inability to cover long-tail risks. Subsequent technological advancements attempted to introduce machine learning and automated classification techniques, aiming to achieve automatic extraction and classification decisions of supplier characteristics through algorithmic models.

[0003] For example, the invention patent CN117216270B uses a bidirectional gated recurrent unit (Bi-GRU) combined with a graph convolutional neural network (GCN) to deeply mine historical text sets of manufacturers and suppliers. It achieves initial supplier matching by establishing semantic relationships between texts. However, a deeper analysis of its underlying logic reveals that the solution relies too heavily on the representation of the "textual semantic" dimension. In the actual context of supply chain management, core structured numerical features such as the supplier's financial robustness, on-time delivery rate, and quality fluctuation indicators often carry more critical performance evaluation information. A single textual feature dimension is insufficient to construct a complete digital profile of suppliers, leading to evaluation biases when performing refined classification and grading tasks due to the lack of key numerical features.

[0004] Existing technologies for addressing supplier classification and grading suffer from two main problems. First, there's an imbalance between the breadth of feature extraction and the depth of data fusion. Current fusion mechanisms often limit themselves to simple linear superposition at the feature vector level, failing to address the fundamental mathematical distinction between structured numerical indicators and unstructured semantic information in the feature space. This leads to the disruption of the coupling relationships between multi-source features. Second, existing classification standards lack dynamic adaptability, making it difficult to flexibly adjust weights according to the dynamic needs of different industry backgrounds and management cycles. The "feature silos" phenomenon and the "static evaluation logic" directly result in one-sided classification results, failing to meet the urgent needs of modern enterprises for lean, multi-dimensional, lifecycle management of suppliers.

[0005] How to build an automated hierarchical classification framework that can deeply integrate multi-source heterogeneous features, accurately depict the comprehensive strength of suppliers, and has a high degree of dynamic adaptability, while ensuring the universality of data acquisition, has become a technical bottleneck that urgently needs to be overcome in the field of supply chain informatization. Summary of the Invention

[0006] The purpose of this invention is to provide a supplier classification method based on multi-dimensional feature fusion, which solves the technical problems in the prior art, such as the difficulty in deep fusion of multi-source heterogeneous features, the semantic difference in feature space between structured numerical and unstructured text information, and the lack of dynamic adaptability of classification standards, which lead to insufficient accuracy in supplier evaluation and limited decision-making flexibility.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] The supplier classification method based on multi-dimensional feature fusion includes the following steps:

[0009] Step 1: Obtain the original dataset of the target supplier. The original dataset includes structured business indicator data, unstructured text semantic data, and historical time series data. The historical time series data covers the supplier's various structured indicators and their corresponding unstructured text records for at least 36 months.

[0010] Step 2: Perform normalization preprocessing on the structured business indicator data;

[0011] Step 3: Input the normalized structured business indicator data into the structured feature mapping network to obtain the structured feature vector; input the unstructured text semantic data into the unstructured semantic representation network to obtain the unstructured semantic feature vector.

[0012] Step 4: Map the structured feature vectors to the same dimension as the unstructured semantic feature vectors through a linear projection layer to obtain aligned structured feature vectors;

[0013] Step 5: Employ a cross-modal attention mechanism, using the aligned structured feature vector as the query and the unstructured semantic feature vector as the key and value, to perform feature fusion and obtain the interactive feature vector;

[0014] Step 6: Perform time series modeling on the interaction feature vector sequence of the supplier over the past 36 months to obtain the time series comprehensive feature vector;

[0015] Step 7: Input the temporal integrated feature vector into a deep hierarchical classifier based on multi-task learning, and output the category and grade of the supplier; wherein, the training process of the deep hierarchical classifier includes an ordinal regression loss function and a contrastive learning loss function;

[0016] Step 8: Based on the dimension priority coefficients set by the administrator, dynamically adjust the attention weights of the cross-modal attention mechanism to achieve dynamic adaptation of the hierarchical results;

[0017] Step 9: Use the closed-loop feedback optimization module to incrementally learn and fine-tune the model based on the supplier's actual business performance data.

[0018] Furthermore, the normalization preprocessing in step 2 employs an extreme value normalization algorithm to map each structured index to... Interval.

[0019] Furthermore, the structured feature mapping network is a multilayer perceptron with residual connections, which includes an input layer, three hidden layers with residual connections, and an output layer. The number of neurons in each hidden layer is set in a ratio of 128, 256, and 128, and a linear rectified function is used as the activation function.

[0020] Furthermore, the unstructured semantic representation network is an encoder based on a pre-trained Transformer structure, which has 12 hidden layers, 16 heads for the multi-head attention mechanism, and a word embedding dimension of 768.

[0021] The encoder transforms the input text sequence into a high-dimensional dense semantic vector, and then obtains an unstructured semantic feature vector through a global average pooling layer and a linear projection layer.

[0022] Furthermore, the cross-modal attention mechanism described in step 5 calculates the interaction feature vector using the following formula. :

[0023] ;

[0024] in, , , All are learnable weight matrices; The aligned structured feature vectors; For unstructured semantic feature vectors; superscript Indicates matrix transpose; The dimension of the key vector is the same as the dimension of the query vector; It is a normalized exponential function; This is the fused interactive feature vector.

[0025] Furthermore, the time series modeling described in step 6 employs a bidirectional long short-term memory network combined with a time attention mechanism to model the interaction feature vector sequence of the supplier over the past 36 months, generating a time-series comprehensive feature vector.

[0026] Furthermore, the ordinal regression loss function described in step 7 is implemented using a cumulative link model, specifically as follows:

[0027] Assuming there are multiple levels One, model output scalar value Through a set of learnable thresholds Will Probabilities mapped to each level:

[0028] ;

[0029] in, Represents probability; This indicates the supplier's true rating. For hierarchical indexing; For the sigmoid function, ; For the first A learnable threshold parameter; This represents the scalar value output by the model; thus, it determines whether the sample belongs to the th... The probability of each level is:

[0030] ;

[0031] definition , Ordinal regression loss function takes the negative log-likelihood:

[0032] ;

[0033] in, For ordinal regression loss; The total number of training samples; For sample index; For the first The true rank of each sample; For the first The grade value corresponding to each sample; It represents the natural logarithm.

[0034] Furthermore, the contrastive learning loss function described in step 7 is:

[0035] ;

[0036] in, To compare the losses; Represents the interaction feature vector of the current sample; This represents the set of positive samples at the same level as the current sample. Summing over all positive samples in the set; This represents the set of negative samples at different levels from the current sample. Summing over all negative samples in the set; The cosine similarity function; For temperature parameters; Represented by natural constant An exponential function with base 0; It represents the natural logarithm.

[0037] Furthermore, the dynamic adjustment of attention weights in the cross-modal attention mechanism described in step 8 specifically includes:

[0038] The vector is constructed by receiving priority coefficients for each dimension set by the administrator through the weight biasing interface. ,in The number of dimensions;

[0039] Will An attention bias vector is generated by mapping to the same dimension as the query vector through a linear layer. ;

[0040] In cross-modal attention calculation, the attention score is modified as follows:

[0041] ;

[0042] in, This is the modified attention score matrix; , , , , The definition is the same as that in claim 5; The attention bias vector is generated from the priority coefficients; then... Perform Softmax normalization to obtain the adjusted attention weights.

[0043] Furthermore, the incremental learning fine-tuning described in step 9 employs a gradient update strategy with memory replay, extracting representative old samples from the historical database and mixing them with new samples to participate in gradient calculation, in order to prevent catastrophic forgetting.

[0044] Furthermore, the cross-modal attention mechanism described in step 5 is implemented using multi-head attention, assuming the number of attention heads is 1. The key vector dimension of each head is Then the total key dimension ;

[0045] For each head Using a learnable weight matrix , , Calculate query ,key ,value ;

[0046] Then calculate the output of the head. ;

[0047] The outputs of all heads are concatenated and then passed through a linear layer. Fusion yields interactive feature vectors. ,in This indicates a splicing operation.

[0048] Furthermore, in the ordinal regression loss function, the number of ranks... For integers greater than or equal to 3, the learnable threshold Through reparameterization Guarantee strict incrementality, among which These are learnable parameters;

[0049] Probability calculations use the cumulative distribution function of the normal distribution. Replace the sigmoid function and introduce a scale parameter. The sample belongs to the level The probability is defined as ,in , .

[0050] Furthermore, the temperature parameter in the contrastive learning loss function The value ranges from 0.05 to 0.5. Positive sample pairs consist of suppliers of the same level, while negative sample pairs consist of suppliers of different levels.

[0051] Furthermore, the total loss function for model training is a weighted sum of ordinal regression loss and contrastive learning loss, with the weights dynamically adjusted using an uncertainty-weighted method.

[0052] Let the noise parameters corresponding to ordinal regression loss and contrastive learning loss be respectively , The total loss is ,in , For learnable parameters, It represents the natural logarithm.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] This invention constructs a dual-path parallel heterogeneous feature extraction architecture to deeply mine structured business indicators and unstructured text semantics separately, and employs a cross-modal attention mechanism to achieve deep fusion of multi-source heterogeneous features. Compared to traditional methods that rely on single-dimensional features or simple linear superposition, this invention can construct a complete digital profile of suppliers from multiple core evaluation domains such as financial robustness, delivery efficiency, quality consistency, and technological innovation, effectively avoiding evaluation misjudgments caused by missing information dimensions. The structured feature mapping network captures the deep coupling relationship between indicators through residual connections and multi-layer nonlinear transformations; the unstructured semantic representation network accurately understands the specific contextual meaning of supply chain professional terms through domain-adaptive pre-training; and the cross-modal attention mechanism enables interactive enhancement between numerical indicators and textual information, significantly improving the comprehensiveness and accuracy of feature representation.

[0055] This invention introduces a joint optimization strategy based on ordinal regression loss and contrastive learning constraints in the hierarchical decision-making stage. Ordinal regression loss models the ordered relationship between levels through a learnable threshold, ensuring the logical coherence of the level division. Contrastive learning constraints bring samples of the same level closer together and push samples of different levels further apart in the feature space, constructing a more discriminative classification boundary. The synergistic effect of these two strategies enables the model to maintain robust classification performance even when dealing with real-world supply chain scenarios with extremely uneven sample distributions, significantly improving its ability to identify high-risk suppliers. The time-series modeling module further captures the dynamic evolution of supplier performance, combining historical trends with the current state, making the hierarchical results more reflective of the true development trend of suppliers.

[0056] This invention's dynamic weight adaptation mechanism enables managers to adjust the priority of each evaluation dimension in real time based on corporate strategic adjustments or industry cycle changes. It directly intervenes in the cross-modal fusion process through attention bias vectors, achieving smooth switching of evaluation preferences without retraining the model. This feature allows the system to accurately adapt to the personalized procurement management needs of companies in different industries and at different stages of development, flexibly configurable from risk-sensitive to cost-sensitive models. The closed-loop feedback optimization module, through incremental learning and memory replay mechanisms, continuously fine-tunes the model using actual business performance data, allowing the grading standards to evolve with the supply chain environment, effectively preventing model aging and concept drift.

[0057] The automated hierarchical classification framework constructed in this invention achieves intelligent processing throughout the entire process, from data collection, feature extraction, dynamic modeling to hierarchical decision-making, significantly reducing the subjectivity and lag of traditional assessments that rely on human experience. By storing high-dimensional supplier features in a vector database, it supports millisecond-level similar supplier retrieval and clustering analysis, providing real-time decision support for business scenarios such as enterprise risk warning, supplier access, and classification management. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0059] Figure 1 This is a simplified flowchart of the method described in this invention.

[0060] Figure 2 This is a flowchart of the method described in this invention.

[0061] Figure 3 This is one of the system operation interface diagrams of the present invention.

[0062] Figure 4 This is the second diagram of the system operation interface of the present invention. Detailed Implementation

[0063] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0064] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0065] In the supplier classification method based on multi-dimensional feature fusion provided by this invention, the system operation logic relies on a highly integrated and modular deep learning architecture during the specific implementation process.

[0066] The deep learning architecture was designed from the outset to fully consider the complexity, heterogeneity, and dynamic characteristics of global supply chain data, and achieves full capture of supplier lifecycle data through a multi-source data acquisition engine.

[0067] In practical implementation, the multi-source data acquisition engine not only connects to the enterprise's internal ERP (Enterprise Resource Planning) system, SRM (Supplier Relationship Management) system, and MES (Manufacturing Execution System), but also connects to external commercial databases, government information disclosure systems, mainstream industry media, and third-party credit rating agencies through standardized API interfaces. Simultaneously, by periodically extracting historical business records from suppliers through a data warehouse, it constructs monthly snapshots of structured indicators for the past 36 months, along with corresponding historical versions of unstructured documents, providing a complete data foundation for time series modeling. This comprehensive data access approach ensures the diversity of input data, providing a solid data foundation for subsequent deep feature extraction.

[0068] In terms of structured business indicator data processing, supplier performance is refined into four core evaluation domains. The financial stability dimension, serving as the cornerstone for assessing a supplier's long-term viability, includes not only basic debt-to-equity ratio and quick ratio, but also in-depth financial indicators such as cash flow turnover days, working capital adequacy ratio, and net profit growth rate over the past three years. The delivery efficiency dimension focuses on the supplier's physical response speed and execution capabilities. It constructs a quantitative description of the supplier's logistics support capabilities by collecting dynamic data in real time, such as on-time order delivery rate, average lead time deviation, capacity utilization rate, and emergency order processing cycle. The quality consistency dimension is the core of the evaluation, encompassing indicators such as incoming material qualification rate, production process capability index (Cpk), after-sales failure rate, and the effectiveness of quality system certification. These indicators collectively constitute a digital mirror of the supplier's quality performance. The technological innovation dimension evaluates the supplier's potential competitiveness and collaborative innovation capabilities in technological evolution through R&D investment ratio, number of patents, new product contribution rate, and the professional qualifications distribution of the technical team.

[0069] For the aforementioned structured business indicator data, this invention implements a rigorous engineering preprocessing procedure. Considering the significant differences in the dimensions between different indicators, such as the number of patents usually being an integer while the material acceptance rate is a percentage, this invention utilizes an extreme value normalization algorithm to map each indicator to the [0, 1] interval based on its historical extreme values ​​over the past five fiscal years.

[0070] For unstructured text semantic data, the processing method of this invention demonstrates a deep application of natural language processing technology. A web crawler deployed in a distributed crawling system can monitor industry news and third-party credit rating texts at a preset frequency, while annual audit reports and compliance assessment reports from the internal office automation system are converted into standardized text streams. This text data contains a wealth of soft information that is difficult to capture with structured indicators, such as corporate business philosophy, potential legal risks, industry reputation, and management changes.

[0071] In the feature extraction stage, the dual-path parallel heterogeneous feature extraction architecture constructed in this invention plays a core role. The first path, a structured feature mapping network, is specifically designed for nonlinear feature mining of the aforementioned quantification metrics. The network employs a multilayer perceptron (MLP) structure, with its input layer dimension corresponding to the total number of structured metrics.

[0072] To address the vanishing gradient problem in deep network training, this invention introduces residual connections between the three hidden layers, meaning that the output of each layer is element-wise added to its input. The number of neurons in each hidden layer has been optimized through engineering, set to 128, 256, and 128 respectively. This "hourglass-shaped" structure helps to extract deep coupling relationships between metrics.

[0073] Meanwhile, the use of the Rectified Linear Unit (ReLU) as an activation function enhances the model's ability to capture nonlinear patterns, ultimately resulting in a more structured feature vector output. It encapsulates all the quantitative operational characteristics of the supplier.

[0074] The second unstructured semantic representation network employs a pre-trained Transformer encoder, a choice based on the Transformer's superior performance in handling dependencies in long texts. This encoder features 12 deep hidden layers, 16 heads for the multi-head attention mechanism, and a word embedding dimension of 768.

[0075] In actual operation, the input text sequence first undergoes word segmentation and positional encoding. Then, under the multi-head attention mechanism, the model can identify key warning words (such as "default," "litigation," and "equity freeze") and their contextual semantics. After 12 layers of complex nonlinear transformations, the text sequence is converted into a high-dimensional dense semantic vector. To further compress information and preserve global semantics, this invention then uses a global average pooling layer and a linear projection layer to obtain an unstructured semantic feature vector. .

[0076] The core of this invention lies in the introduction of a heterogeneous feature alignment and deep fusion mechanism.

[0077] In traditional supplier evaluation, structured data and unstructured text are often processed separately or simply weighted and concatenated, ignoring the semantic relationship between the two. Because structured feature vectors... Derived from numerical space, rather than structured semantic feature vectors Originating from the text manifold space, the two naturally exhibit dimensional mismatch and semantic differences.

[0078] To address this issue, the present invention first uses a linear projection layer to transform the structured feature vector... Mapping to unstructured semantic feature vectors For the same dimensions, dimension alignment is achieved. Subsequently, deep interaction and alignment of features are achieved through the cross-modal attention mechanism in feature fusion layer 4. Specifically, this invention aligns the structured feature vectors... As a query vector, it is mapped to the interaction space through a series of linear transformations, while simultaneously transforming the unstructured semantic feature vector. This is mapped to a key vector and a value vector. Through calculation... and The dot product similarity is calculated and then normalized using Softmax to obtain an attention weight distribution. This distribution indicates which unstructured text information is most worthy of attention given the current supplier's structured metrics performance.

[0079] For example, when a supplier's financial indicators show abnormal fluctuations, the attention mechanism automatically increases the weight of the cash flow description section in the audit report. The fused interaction feature vector. This achieves deep coupling of information from two heterogeneous spaces, providing higher-order feature inputs for subsequent hierarchical decision-making. The formula for calculating cross-modal attention is:

[0080] ;

[0081] in, , , All are learnable weight matrices; The aligned structured feature vectors; For unstructured semantic feature vectors; superscript Indicates matrix transpose; The dimension of the key vector (i.e. The dimension (the same as the query vector dimension) is used to scale the dot product result to prevent gradient vanishing. The normalized exponential function transforms the elements of the input vector into a probability distribution. This is the fused interactive feature vector.

[0082] Considering that supplier performance is a dynamic process that evolves over time, this invention incorporates a time-series modeling module after feature fusion layer 4. This module employs a bidirectional long short-term memory (Bi-LSTM) network, capable of simultaneously capturing both the forward growth trend and backward historical constraints of supplier performance. By modeling the fused feature vector sequence of suppliers over the past 36 months, Bi-LSTM can identify cyclical operating patterns or sudden breakpoints.

[0083] In practical applications, for cases where unstructured text data is missing for some months, the system will adopt a nearest neighbor interpolation strategy: if a certain month only has structured indicators but no corresponding unstructured text records, the average of the unstructured semantic feature vectors of the months before and after using the supplier will be used as the approximate feature of the current month, or the global text feature cluster center will be used to fill the gap, so as to ensure the continuity of the time series.

[0084] By incorporating a time attention mechanism, this module assigns unequal contribution levels to different historical time steps, with recent performance typically given higher weights, while longer-term performance is primarily used to assist in determining stability. The final generated temporal comprehensive feature vector... It achieves a full-dimensional digital abstraction of suppliers' "past-present-future".

[0085] In the hierarchical decision-making stage, the hierarchical decision-maker adopts a deep hierarchical classification architecture based on multi-task learning. This architecture achieves collaborative optimization of classification and hierarchical classification.

[0086] The first subtask header, "Category Prediction Header," maps suppliers to their respective business roles, such as manufacturing suppliers, service suppliers, etc. This is achieved through a Softmax regression function, outputting a probability distribution for each category. The second subtask header, "Rank Assessment Header," then performs refined rank classification within the defined category framework.

[0087] To reflect the inherent order relationship between grades (e.g., grade A is better than grade B), this invention introduces an ordinal regression loss function instead of the traditional cross-entropy loss.

[0088] For example, in practical implementation, the ordinal regression loss is implemented using a cumulative link model:

[0089] Suppose there are K levels in total (K=4, corresponding to A, B, C, and D), and the model outputs a scalar value. Through a set of learnable thresholds Will The probability mapped to each level.

[0090] Specifically:

[0091] ;

[0092] in, Represents probability; This indicates the supplier's true rating. For hierarchical indexing; The sigmoid function is defined as follows: ; For the first A learnable threshold parameter; This is the scalar value output by the model. This allows us to determine which sample belongs to the [number]th [order]. The probability of each level is:

[0093] ;

[0094] definition , The loss function takes the negative log-likelihood:

[0095] ;

[0096] in, For ordinal regression loss; The total number of training samples; For sample index; For the first The true rank of each sample; For the first The grade value corresponding to each sample; It represents the natural logarithm.

[0097] To further enhance the model's robustness when handling small-scale samples or highly imbalanced class distributions, a contrastive learning constraint is introduced during model training. In real-world supply chain scenarios, the number of Grade A suppliers (high-quality suppliers) is often far less than that of Grade C or D suppliers. Contrastive learning, by defining a contrastive loss term in the feature space, forces the model to narrow the feature distance between suppliers of the same grade while widening the feature distance between suppliers of different grades. The mathematical expression for the contrastive loss function is:

[0098] ;

[0099] in, To compare the losses; Represents the interaction feature vector of the current sample; This represents the interaction feature vector of positive samples at the same level as the current sample, with the summation sign... Sum all positive samples in the current training batch; This represents the interaction feature vector of negative samples at different levels from the current sample, with the summation sign... Sum all negative samples in the current training batch; The cosine similarity function; For temperature parameters; Represented by natural constant An exponential function with base 0; It represents the natural logarithm. A more discriminative manifold structure is constructed within the feature space, significantly improving the clarity of classification boundaries in complex noisy environments.

[0100] Furthermore, this invention integrates a dynamic weight adaptation algorithm, providing enterprises with extremely high flexibility in management decision-making. Through a weight adjustment interface, procurement managers can set priority coefficients for each evaluation dimension (financial, delivery, quality, and technical), forming a vector. It can be expanded according to the actual number of dimensions. An attention bias vector is generated by mapping to the same dimension as the query vector in cross-modal attention through a linear layer. In the cross-modal attention computation process, the bias vector is... Adding this to the dot product of the query and the key, the modified attention score is:

[0101] ;

[0102] in, This is the modified attention score matrix; , , , , The definition is the same as in the cross-modal attention formula; This is the attention bias vector generated from the priority coefficients. Then... Perform Softmax normalization.

[0103] because By influencing the attention weights of different location features, the model's attention to different structured indicators and their corresponding text information can be adjusted, thereby achieving a smooth switch of evaluation preferences from risk-sensitive to cost-sensitive, and the entire process does not require retraining the model.

[0104] The closed-loop feedback optimization module of this invention ensures the system's self-evolution capability. This module monitors the actual business performance after the implementation of hierarchical decisions in real time, such as supplier defaults and delivery quality deviations during actual contract fulfillment. Business outputs are transformed into "real labels" and stored in a data training pool. Through incremental learning technology, the model uses these newly generated labeled data for rapid online fine-tuning, enabling the model parameters to continuously adapt to new supply chain environments and industry trends.

[0105] At the physical implementation level, this invention operates within a supply chain brain center comprised of a high-performance computing server cluster. This cluster is equipped with a top-tier GPU array, achieving 100 TFLOPS of single-precision floating-point performance, ensuring real-time performance during large-scale Transformer network training and multi-dimensional feature alignment. Simultaneously, it employs an architecture combining a distributed database and a vector database, with the vector database specifically used to store encoded high-dimensional feature vectors of suppliers. It supports millisecond-level similar supplier retrieval, enabling rapid identification of groups with performance similar to the target supplier, providing a reference for risk warning.

[0106] To further verify the practical effect of the supplier classification method based on multi-dimensional feature fusion provided by the present invention, specific embodiments and comparative data are provided below.

[0107] Example 1: In an example for a large electronics manufacturing company, this example is applied to manage its 8,500 suppliers worldwide. The dataset contains complete business records from the past five years, including 32 structured metrics, and 120,000 unstructured audit reports and public opinion texts. The model was trained using the Adam optimizer with a learning rate of 1e-4 and a batch size of 64. Temperature parameters were compared during the learning process. Set to 0.07.

[0108] Comparative Example 1: A traditional supplier evaluation system based on the Weighted Scoring Model was adopted. This system relies solely on the linear weighting of structured indicators, and the weights are set by expert experience.

[0109] Table 1 shows the performance comparison data of Example 1 and Comparative Example 1 on the same test set:

[0110] Table 1:

[0111]

[0112] As can be seen from the data tables above, this invention demonstrates an overwhelming advantage in core metrics such as classification accuracy and F1-score. Particularly in terms of the detection rate of risky suppliers, this invention, through deep integration of unstructured text features (such as public opinion warnings) and time-series modeling, increases the detection rate from 54.1% to 87.5%, meaning that a large number of potential supply disruption risks are accurately identified before they materialize. Although the individual evaluation time of this invention is slightly increased (185ms), it is still far below the millisecond-level real-time requirement and does not affect the performance of large-scale online evaluation at all.

[0113] In a further preferred embodiment, each residual block of the structured feature mapping network is finely designed. Each residual block contains two fully connected layers, followed by a layer normalization operation. The introduction of layer normalization effectively alleviates the model instability caused by drastic fluctuations in the input feature distribution, resulting in a training convergence speed improvement of over 30%. Simultaneously, at the output of the MLP, this invention adds a feature projection layer to project the structured feature vectors... The dimensions are forced to align with the unstructured semantic feature vectors. The same dimension provides a more standardized input for subsequent cross-modal dot product operations mathematically.

[0114] For unstructured semantic representation networks, this invention introduces a vocabulary expansion specific to the industrial supply chain domain, building upon a pre-trained Transformer. Through secondary pre-training on over 500GB of industrial corpora, trade regulations, and supply chain contract texts, the encoder can accurately understand the deeper meanings of specialized terms such as "back-to-back payment," "force majeure exemption," and "yield ramp-up" in specific contexts. This injection of domain knowledge enhances the semantic feature vectors... It possesses extremely high semantic parsing accuracy, significantly reducing evaluation bias caused by general word segmenters misunderstanding technical terms.

[0115] In the cross-modal attention calculation of feature fusion layer 4, a linear attention mechanism is introduced to further improve computational efficiency. By utilizing the associative law of matrix multiplication, the computational complexity is reduced from the quadratic order of the sequence length to the linear order. This allows the system to maintain extremely high throughput even when processing annual audit reports of tens of thousands of words. Simultaneously, this invention introduces an additional "global context bias term" when calculating attention weights. This term is composed of the supplier's historical rating distribution information, thus implicitly introducing prior knowledge into the current feature fusion process.

[0116] The Bi-LSTM network in the time series modeling module employs a three-layer stacked structure. Each LSTM layer has 512 hidden units, and a dropout rate of 0.3 is applied to prevent overfitting. In the time attention mechanism, a weighted window based on a Gaussian kernel function is designed to allow the model to handle noisy data in the time series more smoothly. For newly added suppliers with less than 36 months of historical data, the module automatically initiates a "cold start strategy," using similar supplier feature transfer technology to fill in missing time steps with cluster center features, ensuring the consistency of evaluation results.

[0117] The hierarchical decision-maker employs a dynamic loss weighting balancing algorithm for multi-task learning. Based on the homoscedasticity uncertainty of each of the two sub-tasks (class prediction and class assessment) during training, the algorithm automatically adjusts their weights in the total loss function. When the class assessment task struggles to converge due to sparse labels, the algorithm automatically increases its weight, forcing the model to favor that task. Furthermore, the ordinal regression loss function, implemented with a set of learnable thresholds, divides continuous predictions into discrete class intervals, ensuring automatic tuning of the class decision boundaries.

[0118] The closed-loop feedback optimization module employs a gradient update strategy with a "memory replay" mechanism during incremental learning. To prevent "catastrophic forgetting" (i.e., forgetting features of old samples when learning new samples) during training on new data, the module extracts a small subset of representative old samples (support vectors) from the historical database and mixes them with the new samples to participate in gradient calculation. This strategy ensures the logical consistency and stability of the grading criteria over a long time span.

[0119] The server cluster involved in this invention is deployed using containerization technology and managed uniformly through Kubernetes. Each microservice node dynamically scales according to its load; for example, during large-scale vendor re-evaluation at the end of a quarter, the system automatically allocates more computing nodes to participate in Transformer encoding computation. The vector database uses a high-performance index structure based on the HNSW algorithm, supporting millisecond-level K-nearest neighbor (KNN) searches within a feature vector space of tens of millions.

[0120] This invention provides a supplier classification and grading method based on multi-dimensional feature fusion. Through the synergistic effect of the aforementioned modules, it achieves a highly automated and intelligent closed-loop supplier management system. It can not only process massive amounts of heterogeneous raw supply chain data, but also uncover hidden operational risks and collaborative value behind the data through a deep neural network architecture. Its implementation significantly reduces the labor costs for enterprises in the supplier onboarding, evaluation, and grading processes, and further enhances the resilience of the supply chain through scientific algorithmic logic. This enables enterprises to possess a data-driven, accurate, and reliable decision support system when facing global competition and an uncertain environment.

[0121] This invention maps the feature vectors of different suppliers to the same manifold space, enabling the system to automatically identify supplier groups with similar behavioral patterns or shared potential risks (such as common upstream raw material suppliers). This "group profiling" capability provides enterprises with the possibility of transforming their supply chain security strategies from "individual defense" to "network-based collaborative defense."

[0122] Regarding data security and privacy protection, this invention strictly adheres to data anonymization protocols during implementation. All sensitive text involving supplier trade secrets undergoes anonymization processing based on named entity recognition before being input into the unstructured semantic representation network, ensuring the protection of supplier privacy information without affecting semantic understanding.

[0123] Example 2: To further optimize the deep fusion effect of heterogeneous features, this example extends the cross-modal attention mechanism to multiple heads based on Example 1 and provides detailed network parameter configuration.

[0124] In this embodiment, the structured feature mapping network adopts a multilayer perceptron (MLP) structure, and its input layer dimension corresponds to the total number of structured indicators (32 dimensions in this embodiment). The network contains four hidden layers with the following numbers of neurons: 128, 256, 256, and 128, respectively. Each hidden layer is followed by a batch normalization layer and a linear rectified function.

[0125] To address the vanishing gradient problem in deep networks, a residual connection is introduced between the second and third hidden layers. This means the output of the third layer is the sum of the transformation result of the third layer and the output of the second layer. The network output layer dimension is set to 256, resulting in a structured feature vector. .

[0126] The domain-adaptive pre-training of unstructured semantic representation networks is as follows:

[0127] The unstructured semantic representation network is based on the Transformer-based BERT model and is pre-trained on more than 1TB of supply chain domain corpus (including purchase contracts, audit reports, industry standards, laws and regulations, etc.) to enable the model to accurately understand professional terms such as "payment period", "obsolete inventory" and "VMI mode".

[0128] In this embodiment, the Transformer encoder has 12 layers, the multi-head attention heads are 12, and the hidden layer dimension is 768. The input text, after being segmented and padded to a fixed length of 512, is input to the encoder. The output vector marked with [CLS] is taken as the initial representation of the text, and then mapped to 256 dimensions through a linear projection layer to obtain the unstructured semantic feature vector. .

[0129] Feature alignment and multi-head cross-modal attention fusion are as follows:

[0130] because and The output layer dimensions have already been aligned (all 256 dimensions), so no additional linear projection layer is needed. This embodiment directly outputs the aligned dimensions... As a query vector As key and value vectors, they are fused using a multi-head attention mechanism.

[0131] In this embodiment, the number of attention heads is set to The key dimension of each head For each head ( ), using learnable weight matrices respectively , , Project the input into the query, key, and value space:

[0132] ;

[0133] in , , The first The query, key, and value vectors of each head; , , All are learnable weight matrices; These are structured feature vectors; This is an unstructured semantic feature vector.

[0134] Then calculate the attention output of the head:

[0135] ;

[0136] in For the first The output vector of each head; It is a normalized exponential function; Calculate the similarity matrix between the query and the key; This is a scaling factor used to prevent gradient vanishing; It is a value vector.

[0137] After concatenating the outputs of all the heads, pass them through a linear layer. The final interaction feature vector is obtained by fusion. :

[0138] ;

[0139] in This represents a vector concatenation operation; The weight matrix is ​​a learnable matrix; This is the fused interactive feature vector.

[0140] The multi-head mechanism enables the model to focus on multiple correlation patterns between structured indicators and textual information from different subspaces. For example, one head focuses on financial indicators and risk descriptions in audit reports, while another head focuses on quality indicators and quality complaints in public opinion.

[0141] This embodiment uses the same optimizer (Adam, learning rate 1e-4) and batch size 64 as in Embodiment 1, and trains on the same dataset for 200 epochs.

[0142] To verify the effectiveness of multi-head cross-modal attention, we compared it with the single-head attention in Example 1, and the results are shown in Table 2 below:

[0143] Table 2:

[0144]

[0145] The results show that multi-head cross-modal attention further improves the refinement of feature fusion, especially by 2.8 percentage points in the detection rate of risk suppliers. This embodiment adopts a multi-head cross-modal attention mechanism, which enables the model to capture the complex correlation between structured indicators and unstructured text in parallel from multiple subspaces, solving the problem of single-head attention focusing on a single mode. At the same time, through domain-adaptive pre-training, the semantic representation network can accurately understand supply chain terminology, significantly improving the quality of unstructured features.

[0146] Example 3: This example is a further optimization based on Example 1. In this example, for the ordered relationship between levels and the problem of data imbalance in hierarchical tasks, a detailed implementation of joint optimization of ordinal regression loss and contrastive learning is provided, and an adaptive adjustment method for the weight of the loss function is given.

[0147] Set the number of levels (Corresponding to five levels: A+, A, B, C, and D, to more precisely distinguish high-quality suppliers).

[0148] The model outputs a scalar value in the rating evaluation header of the hierarchical decision maker. Introduce a set of learnable thresholds. The real number line is divided into 5 intervals, corresponding to various levels. The sample belongs to level [number]. The probability is defined as:

[0149] ;

[0150] in This indicates the supplier's true rating. For level index ( ); The cumulative distribution function of the standard normal distribution; For the first A learnable threshold parameter; The scalar value output by the model; Define the learnable scale parameter (initialized to 1.0); and define... , .

[0151] To avoid disordered threshold order, reparameterization is used during training to ensure... ,in For learnable parameters, .

[0152] Ordinal regression loss is the negative log-likelihood:

[0153] ;

[0154] in For ordinal regression loss; The total number of training samples; For sample index; For the first The true rank of each sample; For the first The grade value corresponding to each sample; It represents the natural logarithm.

[0155] During training, positive and negative sample pairs are constructed from each batch. For the current sample... Its interaction feature vector is Other samples of the same level are used as the positive sample set. Different levels of samples are used as negative sample sets. The comparative loss is expressed in the form of InfoNCE:

[0156] ;

[0157] in To compare the losses; Represents the interaction feature vector of the current sample; Indicates sample Interactive feature vectors; To be consistent with the sample A set of positive samples of the same level; To be consistent with the sample Sets of negative samples at different levels; The cosine similarity function; This is the temperature parameter (set to 0.1 in this embodiment). Represented by natural constant An exponential function with base 0; This represents the natural logarithm. This loss forces suppliers of the same level to move closer to each other in the feature space, while suppliers of different levels move further apart, thereby enhancing the discriminative power of the classification boundary.

[0158] The total loss is a weighted sum of the ordinal regression loss and the contrastive learning loss, with the weights dynamically adjusted using an uncertainty-weighted method.

[0159] ;

[0160] in , Here, are learnable parameters, representing estimates of the uncertainty (noise) for the two tasks, respectively; This represents the natural logarithm. The formula originates from uncertainty weighting in multi-task learning, enabling the model to automatically balance the importance of two tasks.

[0161] This embodiment introduces ordinal regression loss into the supplier classification task, uses a normal distribution cumulative function to model the ordered relationship between grades, and introduces reparameterization techniques to ensure the monotonicity of the threshold, solving the problem of traditional classification loss ignoring grade order. At the same time, it combines contrastive learning constraints to bring samples of the same grade closer and push samples of different grades further apart in the feature space, effectively dealing with data imbalance. More importantly, it adopts uncertainty weighted adaptive adjustment of loss weights, avoiding the tediousness of manual parameter tuning, making model training more stable and efficient.

[0162] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0163] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A supplier classification and grading method based on multi-dimensional feature fusion, characterized in that, Includes the following steps: Step 1: Obtain the original dataset of the target supplier. The original dataset includes structured business indicator data, unstructured text semantic data, and historical time series data. The historical time series data covers the supplier's various structured indicators and their corresponding unstructured text records for at least 36 months. Step 2: Perform normalization preprocessing on the structured business indicator data; Step 3: Input the normalized structured business indicator data into the structured feature mapping network to obtain the structured feature vector; input the unstructured text semantic data into the unstructured semantic representation network to obtain the unstructured semantic feature vector. Step 4: Map the structured feature vectors to the same dimension as the unstructured semantic feature vectors through a linear projection layer to obtain aligned structured feature vectors; Step 5: Employ a cross-modal attention mechanism, using the aligned structured feature vector as the query and the unstructured semantic feature vector as the key and value, to perform feature fusion and obtain the interactive feature vector; Step 6: Perform time series modeling on the interaction feature vector sequence of the supplier over the past 36 months to obtain the time series comprehensive feature vector; Step 7: Input the temporal integrated feature vector into a deep hierarchical classifier based on multi-task learning, and output the category and grade of the supplier; wherein, the training process of the deep hierarchical classifier includes an ordinal regression loss function and a contrastive learning loss function; Step 8: Based on the dimension priority coefficients set by the administrator, dynamically adjust the attention weights of the cross-modal attention mechanism to achieve dynamic adaptation of the hierarchical results; Step 9: Use the closed-loop feedback optimization module to incrementally learn and fine-tune the model based on the supplier's actual business performance data.

2. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The normalization preprocessing described in step 2 employs an extreme value normalization algorithm to map each structured index to... Interval.

3. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The structured feature mapping network is a multilayer perceptron with residual connections, which includes an input layer, three hidden layers with residual connections and an output layer. The number of neurons in each hidden layer is set in a ratio of 128, 256 and 128, and a linear rectified function is used as the activation function.

4. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The unstructured semantic representation network is an encoder based on a pre-trained Transformer structure. The encoder has 12 hidden layers, 16 heads for the multi-head attention mechanism, and a word embedding dimension of 768. The encoder transforms the input text sequence into a high-dimensional dense semantic vector, and then obtains an unstructured semantic feature vector through a global average pooling layer and a linear projection layer.

5. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The cross-modal attention mechanism described in step 5 calculates the interaction feature vector using the following formula. : ； in, , , All are learnable weight matrices; The aligned structured feature vectors; For unstructured semantic feature vectors; superscript Indicates matrix transpose; The dimension of the key vector is the same as the dimension of the query vector; It is a normalized exponential function; This is the fused interactive feature vector.

6. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The time series modeling described in step 6 uses a bidirectional long short-term memory network combined with a time attention mechanism to model the interaction feature vector sequence of the supplier over the past 36 months, generating a time-series comprehensive feature vector.

7. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The ordinal regression loss function described in step 7 is implemented using a cumulative link model, specifically as follows: Assuming there are multiple levels One, model output scalar value Through a set of learnable thresholds Will Probabilities mapped to each level: ； in, Represents probability; This indicates the supplier's true rating. For hierarchical indexing; For the sigmoid function, ; For the first A learnable threshold parameter; This represents the scalar value output by the model; thus, it determines whether the sample belongs to the th... The probability of each level is: ； definition , Ordinal regression loss function takes the negative log-likelihood: ； in, For ordinal regression loss; The total number of training samples; For sample index; For the first The true rank of each sample; For the first The grade value corresponding to each sample; It represents the natural logarithm.

8. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The contrastive learning loss function mentioned in step 7 is: ； in, To compare the losses; Represents the interaction feature vector of the current sample; This represents the set of positive samples at the same level as the current sample. Summing over all positive samples in the set; This represents the set of negative samples at different levels from the current sample. Summing over all negative samples in the set; The cosine similarity function; For temperature parameters; Represented by natural constant An exponential function with base 0; It represents the natural logarithm.

9. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, Step 8, which describes dynamically adjusting the attention weights of the cross-modal attention mechanism, specifically includes: The vector is constructed by receiving priority coefficients for each dimension set by the administrator through the weight biasing interface. ,in The number of dimensions; Will An attention bias vector is generated by mapping to the same dimension as the query vector through a linear layer. ; In cross-modal attention calculation, the attention score is modified as follows: ； in, This is the modified attention score matrix; , , , , The definition is the same as that in claim 5; The attention bias vector is generated from the priority coefficients; then... Perform Softmax normalization to obtain the adjusted attention weights.

10. The supplier classification method based on multi-dimensional feature fusion according to claim 1, characterized in that, The incremental learning fine-tuning described in step 9 employs a gradient update strategy with memory replay, extracting representative old samples from the historical database and mixing them with new samples to participate in gradient calculation, in order to prevent catastrophic forgetting.

Images