A multi-dimensional data fusion merchant operating scene classification method and system

Abstract

The application relates to the field of artificial intelligence, and discloses a multi-dimensional data fusion merchant operation scene classification method and system, aiming to solve the problem of coarse classification granularity and poor real-time performance caused by the dependence on single or static data in the prior art. The method comprises the following steps: acquiring four types of heterogeneous data of merchant static attributes, transaction time sequence, text description and geographic space; respectively extracting high-dimensional features through modal exclusive encoders; performing deep fusion through a cross-modal interaction attention mechanism; constructing a merchant relationship graph containing geographic proximity, customer group similarity and customer sharing relationship, and performing context enhancement on the fused features through a graph attention network; and finally inputting a classifier to output a fine-grained operation scene category. The system comprises five modules of data acquisition, feature encoding, interaction fusion, context enhancement and scene classification. The application realizes accurate, dynamic and fine-grained identification of the operation scene of a merchant, and significantly improves the accuracy and timeliness of classification.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method and system for classifying merchant operation scenarios through multi-dimensional data fusion. Background Technology

[0002] With the rapid development of the digital economy and the continuous improvement of payment infrastructure, merchant operational data is characterized by diverse sources, heterogeneous structure, and high frequency of dynamics. In application scenarios such as smart commerce, precision marketing, and risk management, the refined classification of merchant operational scenarios has become a key prerequisite for improving the level of service intelligence.

[0003] Existing technologies typically rely on single dimensions or static data (such as transaction history, geographic location, or industry tags) to coarsely categorize merchants, making it difficult to comprehensively depict their true operational status. For example, classification methods based solely on transaction amount and frequency cannot distinguish between temporary promotions and stable operations, while relying solely on registered industry tags ignores the cross-industry integration of actual business models. Furthermore, issues such as semantic gaps, temporal asynchrony, and confidence differences exist among multi-source data. Directly concatenating or weighted fusion can easily lead to feature redundancy, noise amplification, or obscuring of key patterns, resulting in classification results that lack stability and interpretability.

[0004] The aforementioned technical shortcomings are particularly prominent in high-requirement scenarios such as merchant onboarding verification by payment institutions. Currently, this field mainly relies on two types of solutions: one is purely manual verification, where reviewers subjectively judge merchant photos and materials, which suffers from inherent bottlenecks such as low efficiency, high misjudgment rate, and inability to handle complex scenarios (such as storefronts and interiors having no direct visual connection); the other is rule-based image comparison solutions, which, although attempting semi-automation, suffer from rigid rules, poor adaptability, and a still high misjudgment rate for non-standard scenarios, failing to fundamentally solve the core difficulties of deep fusion of multimodal information and logical reasoning.

[0005] In summary, existing technologies face multiple challenges in multi-dimensional data fusion, such as heterogeneous data sources, poor adaptability to dynamic evolution, and fixed classification boundaries. These challenges make it difficult to achieve accurate, dynamic, and fine-grained perception of merchants' business scenarios, which severely restricts the service quality and decision-making efficiency of downstream applications. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a method and system for classifying merchant business scenarios through multi-dimensional data fusion, which aims to overcome the technical defects of existing technologies that rely on only a single or static data source, resulting in coarse classification granularity, poor real-time performance, and inability to accurately reflect the real multi-dimensional business formats of merchants.

[0007] To address the aforementioned technical issues, this invention provides a multi-dimensional data fusion method for classifying merchant business scenarios. The core of this method lies in constructing a hierarchical and progressive data fusion and context-aware framework.

[0008] First, a modality-specific deep encoder is used to extract high-dimensional features from merchants' static attributes, dynamic transaction sequences, unstructured text, and geospatial environment, ensuring that the inherent characteristics of each data dimension are fully preserved. Next, a cross-modal interactive attention mechanism is introduced to force the feature representations of different data modalities to undergo deep alignment and information complementarity, generating a comprehensive and internally interconnected multi-dimensional fusion feature. Finally, merchants are placed in a global merchant relationship graph, and graph neural network technology is used to propagate and aggregate neighborhood contextual information to perform social and environmental correction and enhancement of the fusion feature, ultimately achieving accurate, dynamic, and fine-grained classification of merchant operating scenarios.

[0009] According to one aspect of the present invention, a method for classifying merchant operation scenarios through multi-dimensional data fusion is provided, comprising the following steps:

[0010] Acquire multi-source heterogeneous merchant data, which includes merchant static attribute data, merchant transaction time series data, merchant text description data, and merchant geospatial data; optionally, it also includes merchant image data.

[0011] Modality-specific feature encoding is performed on the multi-source heterogeneous merchant data to generate static attribute feature vectors, transaction time-series feature vectors, semantic text feature vectors, and geospatial feature vectors; if merchant image data exists, image feature vectors are generated.

[0012] Cross-modal interaction fusion is performed on the static attribute feature vector, transaction time series feature vector, semantic text feature vector and geospatial feature vector to generate a merchant multidimensional fused feature vector.

[0013] A merchant relationship graph is constructed, and the multidimensional fusion feature vector of the merchants is enhanced with a graph propagation algorithm to obtain the final merchant scene representation vector.

[0014] The final merchant scenario representation vector is input into a preset classifier, and the business scenario classification result of the merchant is output.

[0015] As one embodiment of the present invention, the acquisition of multi-source heterogeneous merchant data specifically includes:

[0016] The merchant static attribute data is obtained from the merchant registration database or a third-party data platform. The merchant static attribute data includes the merchant category code, registered capital, years of establishment, chain brand logo, and approved business scope.

[0017] By connecting with the payment system interface, all transaction records of the merchant within a preset time window are collected in real time to form the merchant's transaction time sequence data. Each transaction record includes transaction amount, transaction timestamp, payment method and product summary information.

[0018] By using web crawling technology or application programming interface calls, user reviews, ratings, tags, and business introductions published by merchants themselves are crawled from public comment platforms, social media, and official merchant websites to form the merchant text description data;

[0019] The precise latitude and longitude coordinates of the merchant are obtained through a geographic information system, and the type, quantity and distribution information of surrounding points of interest are collected within a preset radius centered on the coordinates to form the merchant's geospatial data.

[0020] Furthermore, after acquiring the multi-source heterogeneous merchant data, the method also includes a data preprocessing step, specifically including one-hot encoding or entity embedding processing of the merchant static attribute data; aggregating the merchant transaction time series data according to a preset time step, such as one hour, and calculating the total transaction amount, number of transactions, average order value, and high-frequency product category distribution within each time step to form a multi-dimensional transaction time series; performing word segmentation, stop word removal, part-of-speech tagging, and stemming on the merchant text description data, and using a word vector model to convert the text into an initial word embedding sequence; and performing gridding processing on the merchant geospatial data, statistically analyzing the density of different types of interest points within each grid to form a geospatial feature map.

[0021] As one embodiment of the present invention, the processing of the merchant image data includes:

[0022] Image features are extracted using a large visual model, including photo category verification and visual content information;

[0023] By using a large-scale reasoning model, forward logical reasoning and reverse proof are performed to output consistent results;

[0024] The results output module generates a "compliant", "non-compliant", or "manually reviewed" conclusion based on the reasoning results.

[0025] As one embodiment of the present invention, the modality-specific feature encoding of the multi-source heterogeneous merchant data specifically includes:

[0026] The preprocessed static attribute data of merchants is input into a multilayer perceptron network containing at least two fully connected layers and a nonlinear activation function. Through forward propagation, the static attribute feature vector with a fixed dimension is output.

[0027] The multidimensional transaction time series is input into a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network captures the long-term dependencies of the transaction series from both forward and reverse time dimensions, and splices the final hidden states in the two directions to form the transaction time series feature vector.

[0028] The initial word embedding sequence is input into a pre-trained language model based on a transformer architecture. The pre-trained language model captures deep semantic relationships within the text through a multi-head self-attention mechanism and uses the output vector corresponding to the sentence beginning classification label as the semantic text feature vector.

[0029] The geospatial feature map is input into a convolutional neural network containing multiple convolutional layers, pooling layers, and activation functions. Local spatial patterns with translation invariance are extracted through layer-by-layer convolution operations, and the feature map output by the last convolutional layer is subjected to global average pooling to generate the geospatial feature vector.

[0030] Furthermore, the cross-modal interactive fusion of the static attribute feature vector, transaction time-series feature vector, semantic text feature vector, and geospatial feature vector is specifically implemented through a multimodal interactive attention network, the processing flow of which includes:

[0031] The feature vectors of the four modalities are used as input, and the feature vector of one modality is set as the query vector, while the feature vectors of the other three modalities are set as the key vector and value vector set.

[0032] Calculate the dot product similarity between the query vector and each key vector, and convert the similarity score into attention weights using a normalized exponential function;

[0033] The obtained attention weights are used to perform a weighted summation on the corresponding set of value vectors to generate a context-aware feature vector that incorporates information from other modalities and is specific to the current query modality.

[0034] The above calculation process is repeated, using the feature vector of each modality as the query vector, to obtain four modality feature vectors that have been interactively enhanced.

[0035] The four interactive enhancement modal feature vectors are concatenated in dimension and then reduced in dimension through a linear mapping layer to finally output the merchant's multidimensional fused feature vector.

[0036] As one embodiment of the present invention, the construction of the merchant relationship graph and the use of a graph propagation algorithm to perform context enhancement on the multi-dimensional fused feature vector of the merchants specifically include:

[0037] Each merchant is considered as a node in a graph, and the multidimensional fusion feature vector of the merchant is used as the initial attribute of the node.

[0038] Edges between nodes are established based on preset connection rules, which include: Rule 1: If the geographical straight-line distance between two merchants is less than a preset distance threshold, a geographical proximity edge is established between their corresponding nodes; Rule 2: If the cosine similarity of the target customer profiles of two merchants is greater than a preset similarity threshold, a customer group similarity edge is established between their corresponding nodes; Rule 3: If the customer flow overlap of two merchants within a preset time period, i.e., the average proportion of shared customers to their respective total customers, is greater than a preset overlap threshold, a customer sharing edge is established between their corresponding nodes.

[0039] The constructed merchant relationship graph and the initial attributes of all nodes are input into a graph attention network, which updates the node representation through multi-layer information propagation and aggregation operations. In each layer, each node dynamically assigns different importance weights to its neighboring nodes through a self-attention mechanism, and updates its own feature representation by aggregating the feature information of its neighboring nodes based on these weights.

[0040] After propagation through a graph attention network with a preset number of layers, the output node representation is the final merchant scenario representation vector, which incorporates its position and relationship information in the merchant ecosystem network.

[0041] Furthermore, the step of inputting the final merchant scenario representation vector into a preset classifier and outputting the merchant's business scenario classification result specifically includes:

[0042] The final merchant scenario representation vector is input into a classification head consisting of one or more fully connected layers; the last layer of the classification head uses a normalized exponential function as the activation function, and the number of its output nodes is the same as the total number of predefined business scenario categories.

[0043] The output of the normalized exponential function is a probability distribution vector, where each element represents the probability that the merchant belongs to the corresponding business scenario category;

[0044] The business scenario category corresponding to the element with the highest probability value is selected as the final classification result output. The business scenario categories include, but are not limited to, sub-scenarios such as "business casual dining", "community convenience store", "nighttime entertainment bar", "family activity center", and "popular cafe".

[0045] According to another aspect of the present invention, a multi-dimensional data fusion-based merchant operation scenario classification system is provided, the system being configured to perform the aforementioned method, comprising:

[0046] The data acquisition module is used to acquire static attribute data, transaction time series data, text description data, and geospatial data of merchants from merchant registration databases, payment systems, online platforms, and geographic information systems.

[0047] The feature encoding module integrates a multilayer perceptron network, a bidirectional gated recurrent unit network, a pre-trained language model based on a transformer architecture, and a convolutional neural network, which are used to encode the data of the four modalities into static attribute feature vectors, transaction time series feature vectors, semantic text feature vectors, and geospatial feature vectors, respectively. The feature encoding module can also integrate a large visual model and a large inference model for processing merchant image data.

[0048] The interactive fusion module includes a multimodal interactive attention network, which receives the four feature vectors and generates a comprehensive merchant multidimensional fusion feature vector through cross-modal query-key-value attention calculation.

[0049] The context enhancement module is configured to construct a merchant relationship graph based on the geographical proximity, customer group similarity and customer sharing among merchants, and includes a graph attention network for propagating and aggregating the multidimensional fused feature vectors of the merchants on the graph, and outputting a final merchant scene representation vector that incorporates neighborhood context information.

[0050] The scenario classification module includes a fully connected classification head, which receives the final merchant scenario representation vector and outputs the final determination result of the business scenario category to which the merchant belongs.

[0051] In summary, this application includes at least one of the following beneficial technical effects:

[0052] I. Comprehensive and In-Depth Data Analysis: This invention integrates four heterogeneous data sources—static attributes, dynamic transactions, semantic text, and geospatial data—to construct a three-dimensional, all-encompassing view of merchant operations. By designing a dedicated deep learning encoder for each data modality, it can extract deep, high-dimensional features from the raw data that are difficult to access using traditional methods, significantly improving the richness and accuracy of feature representation.

[0053] II. Innovative Cross-Modal Interactive Fusion Mechanism: This invention abandons simple feature concatenation or weighted fusion methods and introduces a cross-modal interactive attention network. This mechanism can dynamically learn the complex nonlinear dependencies between different data modalities, enabling features of one modality to be modulated and enhanced by contextual information from other modalities, thereby generating a unified feature representation that is logically consistent and informationally complementary, solving the problems of information loss and semantic gap in the process of heterogeneous data fusion.

[0054] Third, the introduction of socialized and environmental context correction: This invention creatively constructs a merchant relationship graph and considers the characteristics of individual merchants within the entire merchant ecosystem network. Utilizing graph attention networks, the system can effectively integrate and correct the scene representation of the target merchant by incorporating macro-environmental information such as the operating status of surrounding merchants and customer flow. This ensures that the classification results are not only based on the merchant's own attributes but also consider the micro-business environment in which it operates, greatly improving the scene adaptability and stability of the classification.

[0055] IV. Refined and Dynamic Classification Results: Combining the above-mentioned technical advantages, this invention can output more refined and dynamic business scenario classifications than traditional merchant category codes. It can distinguish different business models under the same category code, such as differentiating "restaurant" as a "business lunch spot" or a "romantic date spot," and can update the scenario classification in real time based on changes in transaction data and user reviews. This provides high-quality and timely decision-making support for downstream applications such as financial risk control, precision marketing, and commercial site selection. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the overall technical solution architecture of a merchant operation scenario classification method and system proposed in this invention, which involves multi-dimensional data fusion.

[0057] Figure 2 This is a schematic diagram illustrating the core principle framework of the cross-modal interactive attention mechanism in this invention;

[0058] Figure 3 This is a flowchart illustrating the modality-specific feature encoding and preprocessing logic of multi-source heterogeneous merchant data in this invention.

[0059] Figure 4 This is a flowchart illustrating the context-enhanced logic process of merchant relationship graph construction and graph attention network propagation in this invention.

[0060] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the cloud in this invention. Detailed Implementation

[0061] This invention provides a method and system for classifying merchant business scenarios using multi-dimensional data fusion, aiming to address the technical shortcomings of existing technologies that rely on single or static data sources, resulting in coarse-grained classification of merchant business scenarios, poor real-time performance, and an inability to accurately reflect the true multi-dimensional business operations of merchants. This method achieves precise, dynamic, and fine-grained classification of merchant business scenarios by constructing a hierarchical and progressive data fusion and context-aware framework.

[0062] Reference Figures 1 to 5Firstly, the merchant operation scenario classification method of this application includes the following steps: S1: acquiring multi-source heterogeneous merchant data; S2: performing modality-specific feature encoding on the multi-source heterogeneous merchant data; S3: performing cross-modal interactive fusion on the feature vectors of each modality; S4: constructing a merchant relationship graph and using a graph propagation algorithm to enhance the context of the fused features; S5: inputting the final merchant scenario representation vector into a preset classifier and outputting the merchant operation scenario classification result.

[0063] The first step is to execute S1, which involves acquiring multi-source heterogeneous merchant data. This multi-source heterogeneous merchant data includes merchant static attribute data, merchant transaction time-series data, merchant text description data, and merchant geospatial data. Specifically, merchant static attribute data is obtained from a merchant registration database or a third-party data platform. This static attribute data includes the merchant category code, registered capital, years of establishment, chain brand identity, and approved business scope.

[0064] By connecting with the payment system interface, all transaction records of merchants within a preset time window are collected in real time to form merchant transaction time sequence data. Each transaction record includes transaction amount, transaction timestamp, payment method and product summary information.

[0065] By using web crawling technology or application programming interface (API) calls, user reviews, ratings, tags, and business introductions published by merchants themselves are scraped from public comment platforms, social media, and merchants' official websites to form merchant text description data.

[0066] The precise latitude and longitude coordinates of merchants are obtained through a geographic information system. Then, within a preset radius centered on these coordinates, information on the type, quantity, and distribution of surrounding points of interest is collected to form the geospatial data of the merchants.

[0067] After acquiring heterogeneous merchant data from multiple sources, a data preprocessing step is performed. One-hot encoding or entity embedding is applied to the merchant static attribute data. Discrete fields such as merchant category codes and chain brand identifiers are encoded using one-hot encoding, while continuous fields such as registered capital and years of establishment are normalized before being input into the entity embedding layer to generate a uniform-dimensional embedding vector.

[0068] Merchant transaction time-series data is aggregated according to a preset time step, such as one hour, and the total transaction amount, number of transactions, average order value, and distribution of high-frequency product categories are calculated within each time step to form a multi-dimensional transaction time series. Merchant text description data is segmented, stop words are removed, parts of speech are tagged, and stemming is performed. Then, a word vector model is used to convert the text into an initial word embedding sequence.

[0069] The geospatial data of merchants is processed into grids, dividing the area where the merchants are located into several grid units of fixed size. The density of different types of points of interest in each grid is counted to form a geospatial feature map.

[0070] After step S1 is completed, step S2 is executed, which involves modality-specific feature encoding of multi-source heterogeneous merchant data to generate static attribute feature vectors, transaction time-series feature vectors, semantic text feature vectors, and geospatial feature vectors.

[0071] Preprocessed merchant static attribute data is input into a multilayer perceptron network containing at least two fully connected layers and nonlinear activation functions. Through forward propagation, a fixed-dimensional static attribute feature vector is output. The first fully connected layer of the multilayer perceptron network maps the input embedding vector to an intermediate dimension, and the second fully connected layer further compresses it to the target dimension. Each layer is followed by a modified linear unit activation function.

[0072] The multidimensional transaction time series is input into a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network captures the long-term dependencies of the transaction series from both forward and reverse time dimensions, and splices the final hidden states of the two directions to form a transaction time series feature vector.

[0073] The bidirectional gated recurrent unit network comprises forward-gated recurrent units and backward-gated recurrent units, which share the same structure but process in opposite directions. Their hidden state dimensions are preset, and the concatenated vector dimension is twice the hidden state dimension. The initial word embedding sequence is input into a pre-trained language model based on a transformer architecture. The pre-trained language model captures deep semantic relationships within the text through a multi-head self-attention mechanism and uses the output vector corresponding to the sentence-first classification label as the semantic text feature vector.

[0074] The pre-trained language model consists of a twelve-layer transformer encoder block, each containing a multi-head self-attention sub-layer and a feedforward neural network sub-layer. A special classification label is added before the input sequence, and the output of this label in the last layer is the semantic text feature vector. The geospatial feature map is input into a convolutional neural network containing multiple convolutional layers, pooling layers, and activation functions. Translation-invariant local spatial patterns are extracted through layer-by-layer convolutional operations, and the feature map output from the last convolutional layer is subjected to global average pooling to generate the geospatial feature vector.

[0075] A convolutional neural network contains three convolutional blocks. Each convolutional block consists of two convolutional layers, a batch normalization layer, and a modified linear unit activation function. The kernel size is 3x3 with a stride of 1 and the padding method is the same. Each convolutional block is followed by a max pooling layer with a pooling window size of 2x2 and a stride of 2.

[0076] After completing the modality-specific feature encoding, S3 is executed, which involves cross-modal interactive fusion of static attribute feature vectors, transaction time series feature vectors, semantic text feature vectors, and geospatial feature vectors to generate merchant multidimensional fused feature vectors.

[0077] This process is implemented using a multimodal interactive attention network. Feature vectors from four modalities are taken as input, with one modality's feature vector designated as the query vector, and the other three modalities' feature vectors serving as a set of key and value vectors. The dot product similarity between the query vector and each key vector is calculated, and the similarity score is converted into attention weights using a normalized exponential function.

[0078] The obtained attention weights are used to perform a weighted summation of the corresponding value vector sets to generate a context-aware feature vector that incorporates information from other modalities and is specific to the current query modality. This calculation process is repeated for each modality's feature vector as a query vector, resulting in four interactively enhanced modal feature vectors. These four interactively enhanced modal feature vectors are then concatenated dimensionally and dimensionality reduced using a linear mapping layer, ultimately outputting a multi-dimensional fused feature vector for the merchant.

[0079] In the multimodal interactive attention network, the query, key, and value projection matrices are all learnable parameters with the same dimension as the input feature vector. The temperature coefficient of the normalized exponential function is set to one, and the linear mapping layer maps the concatenated high-dimensional vector to the preset target dimension.

[0080] After cross-modal interaction fusion is completed, S4 is executed, which involves constructing a merchant relationship graph and using a graph propagation algorithm to enhance the context of the merchant multidimensional fused feature vector, resulting in the final merchant scenario representation vector. Each merchant is treated as a node in the graph, and the merchant multidimensional fused feature vector serves as the initial attribute of the node.

[0081] Edges between nodes are established based on preset connection rules, which include: Rule 1: If the straight-line geographical distance between two merchants is less than a preset distance threshold, a geographical proximity edge is established between their corresponding nodes; Rule 2: If the cosine similarity of the target customer profiles of two merchants is greater than a preset similarity threshold, a customer group similarity edge is established between their corresponding nodes; Rule 3: If the customer flow overlap of two merchants within a preset time period, i.e., the average proportion of shared customers to their respective total customers, is greater than a preset overlap threshold, a customer sharing edge is established between their corresponding nodes.

[0082] Geographic straight-line distance is calculated using the merchant's latitude and longitude coordinates, with a distance threshold of 500 meters; the target customer profile is composed of dimensions such as age, gender, spending power, and active time period, with a cosine similarity threshold of 0.7; the calculation of customer flow overlap is based on anonymized user identifiers in the payment system, with an overlap threshold of 10%.

[0083] The constructed merchant relationship graph and the initial attributes of all nodes are input into a graph attention network. The graph attention network updates the node representation through multiple layers of information propagation and aggregation operations. In each layer, each node dynamically assigns different importance weights to its neighboring nodes through a self-attention mechanism, and updates its own feature representation by aggregating the feature information of its neighboring nodes based on these weights.

[0084] After propagation through a pre-defined number of graph attention network layers, the output node representation is the final merchant scenario representation vector. This vector incorporates its position and relationship information within the merchant ecosystem network. The graph attention network consists of two propagation layers, each containing the attention coefficients of nodes to their neighbors. Calculated using the following formula:

[0085]

[0086] in, and Let be the feature vectors of nodes i and j, respectively; W be the learnable weight matrix; a be the learnable attention vector; and || denotes vector concatenation. Let represent the set of neighbors of node i. The negative slope parameter of LeakyReLU is set to 0.2. Let k represent the transpose of the attention vector a, where k is the set of neighbors to be traversed. The index variables of all nodes in the array. The update characteristics of node i. Calculated using the following formula:

[0087]

[0088] Where σ is the modified linear unit activation function.

[0089] Finally, S5 is executed, which inputs the final merchant scenario representation vector into a pre-defined classifier and outputs the merchant's business scenario classification result. The final merchant scenario representation vector is input into a classification head consisting of one or more fully connected layers. The last layer of the classification head uses a normalized exponential function as the activation function, and the number of its output nodes is the same as the total number of predefined business scenario categories. The output of the normalized exponential function is a probability distribution vector, where each element represents the probability that the merchant belongs to the corresponding business scenario category. The business scenario category corresponding to the element with the highest probability value is selected as the final classification result output. Business scenario categories include, but are not limited to, sub-scenarios such as "business casual dining," "community convenience store," "nighttime entertainment bar," "family activity center," and "trendy cafe." The classification head contains a hidden layer with a dimension half that of the final merchant scenario representation vector, using a modified linear unit as the activation function, and the output layer dimension equal to the total number of predefined categories.

[0090] It is important to note that there are strict data dependencies between the steps in the entire methodology. The raw data acquired in S1 is the sole input source for feature encoding in S2; the four feature vectors output by S2 are the direct inputs for cross-modal fusion in S3; the merchant multidimensional fusion feature vector generated in S3 constitutes the initial node attributes of the graph neural network in S4; and the final merchant scenario representation vector output by S4 is the sole input to the classifier in S5. The absence or error of any step will cause subsequent steps to fail or produce invalid results.

[0091] Furthermore, the technical feature parameters of this application require further explanation, as follows: Regarding data flow, the original multi-source heterogeneous data, after preprocessing, enters four independent encoding channels. After feature extraction within each channel, the four feature vectors are fed into a unified interactive fusion module. The fused single vector serves as the initial state of a graph node, undergoing multiple rounds of message passing and aggregation on the graph structure. The resulting enhanced vector is then fed into the classifier to complete the decision. The entire process achieves an end-to-end mapping from raw data to high-order semantic representation.

[0092] Regarding parameter validation mechanisms, all preset thresholds and hyperparameters are determined through cross-validation during the model training phase. Geographic proximity thresholds, customer group similarity thresholds, and customer overlap thresholds are all determined through grid search on the validation set, selecting the combination that maximizes classification accuracy. Structural hyperparameters such as the number of layers in the graph attention network, the number of layers in the multilayer perceptron, and the depth of the convolutional neural network are also determined through validation set performance. All numerical parameters undergo range checks before being input into the model to ensure they are within a reasonable range; for example, registered capital cannot be negative, and the number of years since establishment cannot exceed the current year.

[0093] In terms of anomaly handling, if a data source is missing, the system employs a specific strategy for compensation. If merchant static attribute data is missing, it is filled using the average attribute values ​​of similar merchants; if transaction time-series data is empty, a zero-based vector is generated as the transaction time-series feature; if text description data is missing, the semantic text feature vector is set to zero; if geospatial data is unavailable, the geospatial feature vector is set to zero. During the graph construction phase, if customer group similarity or customer overlap cannot be calculated, edges are constructed solely based on geographic proximity. When the classifier receives an anomaly input, it outputs a low-confidence probability distribution and triggers a manual review process.

[0094] The method achieves high-precision, fine-grained, and dynamic classification of merchant operation scenarios through the coordinated work of the above five steps. Its core lies in the fact that modality-specific encoding preserves the unique information of each data source, cross-modal interactive fusion establishes deep connections between heterogeneous data, and graph neural networks introduce social and environmental context correction. The combination of these three elements constitutes a robust and efficient classification framework.

[0095] Furthermore, this invention provides a preferred embodiment, specifically targeting the high-value application scenario of merchant onboarding verification for payment institutions. This embodiment describes a method for verifying the authenticity of merchant business operations based on image recognition and inference models. This method can operate independently or serve as a special case of the aforementioned general method for processing "merchant image data." The method specifically includes the following:

[0096] The data input module receives three photos (storefront photo, interior photo, and cashier photo) uploaded by the merchant, along with the merchant's basic information (merchant abbreviation, business license scope text, and MCC code), and performs standardized format processing on the photos.

[0097] The visual processing module, based on a finely tuned qwen-vi-max, takes a standardized photo as input and outputs the photo category verification results (determining whether it is a valid storefront photo, interior photo, or cashier photo) and visual features and content information (such as store sign text, decoration style labels, and product / equipment labels).

[0098] The reasoning and judgment module, based on a finely tuned qwen-max, takes the output of the visual processing module and basic merchant information as input, and includes two parallel sub-modules:

[0099] The forward logic reasoning submodule performs multi-dimensional compliance checks, including consistency of the main business of the storefront photo, matching degree of the merchant's abbreviation (calculating semantic similarity), matching of the business license scope, and matching of the MCC business scope.

[0100] The reverse proof judgment submodule actively seeks counter-evidence that "three photos are not from the same merchant". The types of counter-evidence include contradictions in main business, address, and brand.

[0101] The results output module integrates the reasoning results and generates the final output according to the rules: "Invalid photo, needs to be re-uploaded", "The three photos are inconsistent, the merchant scenario is not compliant", "The three photos are consistent, the merchant scenario is compliant" or "Manual review required".

[0102] In this preferred embodiment:

[0103] The qwen-vi-max model has been fine-tuned using 100,000 sets of labeled merchant scene photos to optimize its ability to recognize specific elements in merchant scenes, improve the accuracy of photo category judgment, and enhance the accuracy of visual feature extraction.

[0104] qwen-max was fine-tuned by using 50,000 sets of labeled data containing basic merchant information, visual features, and consistency results, including 10,000 sets of reverse proof cases. This allows the model to master the rules of forward logical reasoning and the logic of reverse proof judgment, thereby improving the accuracy of multi-dimensional verification and the accuracy of proof of contradiction recognition.

[0105] Specific implementation examples of this preferred embodiment include "conventional retail merchant scenarios", "scenarios where there is no direct correlation between storefront and interior views", "non-compliant scenarios with counter-evidence", and "edge cases (manual review scenarios)". This allows the present invention to construct a complete, multi-layered, and adaptable multi-dimensional data fusion merchant operation scenario classification solution by combining general methods with preferred embodiments. It not only achieves deep fusion and context awareness of multi-source heterogeneous data in terms of technology, but also has good implementability and scalability in terms of engineering. It can effectively support the demand for high-quality merchant profiles in downstream application scenarios such as financial risk control, precision marketing, commercial site selection, and payment review.

[0106] In addition to the method, this invention also provides a multi-dimensional data fusion-based merchant operation scenario classification system. This system is configured to execute the aforementioned method and includes a data acquisition module, a feature encoding module, an interaction fusion module, a context enhancement module, and a scenario classification module.

[0107] The data acquisition module is used to obtain static attribute data, transaction time series data, text description data, and geospatial data of merchants from merchant registration databases, payment systems, online platforms, and geographic information systems.

[0108] This module contains four sub-units: static attribute acquisition unit, transaction data acquisition unit, text data acquisition unit, and geographic data acquisition unit.

[0109] The static attribute acquisition unit connects to the merchant registration database through a structured query language interface and regularly synchronizes the latest merchant information;

[0110] The transaction data acquisition unit connects with the payment system in real time through a secure socket layer encrypted channel to receive transaction events in a streaming manner;

[0111] The text data acquisition unit deploys a distributed web crawler cluster to periodically crawl publicly available merchant information from designated platforms;

[0112] The geographic data acquisition unit calls the geographic information system application programming interface to request merchant coordinates and surrounding points of interest data in batches.

[0113] The feature encoding module integrates a multilayer perceptron network, a bidirectional gated recurrent unit network, a pre-trained language model based on a transformer architecture, and a convolutional neural network, which are used to encode data of four modalities into static attribute feature vectors, transaction time series feature vectors, semantic text feature vectors, and geospatial feature vectors, respectively.

[0114] The feature encoding module contains four parallel encoding sub-modules, each of which independently processes a data modality.

[0115] The static attribute encoding submodule consists of two fully connected layers. The input dimension is the same as the dimension of the preprocessed static attribute vector, and the output dimension is the preset unified feature dimension.

[0116] The transaction timing encoding submodule adopts a bidirectional gated cyclic unit structure, with the number of hidden units set to a preset value, and supports variable-length sequence input;

[0117] The text semantic encoding submodule loads the pre-trained language model weights, freezes the underlying parameters, and only fine-tunes the top-level classification head;

[0118] The geospatial coding submodule adopts a standard convolutional neural network architecture, with a two-dimensional grid feature map as input and a one-dimensional feature vector as output.

[0119] The interactive fusion module includes a multimodal interactive attention network that receives four feature vectors and generates a comprehensive multidimensional fused feature vector for the merchant through cross-modal query-key-value attention computation. This module maintains four sets of learnable projection matrices, used to map features from each modality to a common query, key, and value space. Internally, the module implements four rounds of attention computation, each round using one modality as the query and the other three as key-value pairs. Finally, the four enhanced features are concatenated and dimensionality reduced using a linear layer.

[0120] The context enhancement module is configured to construct a merchant relationship graph based on the geographical proximity, customer group similarity, and customer sharing among merchants, and includes a graph attention network for propagating and aggregating the multidimensional fused feature vectors of merchants on the graph, outputting a final merchant scene representation vector that incorporates neighborhood context information.

[0121] This module includes a graph construction unit and a graph propagation unit. The graph construction unit maintains a merchant coordinate index, a customer profile database, and a customer flow record table, and can quickly calculate the three types of connection strengths between any two merchants; the graph propagation unit implements the forward computation of the graph attention network, and supports batch node updates and sparse adjacency matrix operations.

[0122] The scene classification module includes a fully connected classification head, which receives the final merchant scene representation vector and outputs the final determination result of the business scene category to which the merchant belongs. This module consists of a hidden layer and an output layer. The dimension of the hidden layer is half the dimension of the input vector, and the dimension of the output layer is equal to the total number of predefined categories. The activation function of the output layer is a normalized exponential function.

[0123] The system's modules communicate with each other through standardized data interfaces. The data acquisition module outputs raw data to a preprocessing buffer; the feature encoding module reads preprocessed data from the buffer and outputs feature vectors to a fusion queue; the interactive fusion module consumes the quadruple features from the fusion queue and outputs fusion vectors to a graph node attribute table; the context enhancement module reads the node attribute table and graph structure data and outputs enhanced representation vectors to a classification input queue; the scene classification module consumes the classification input queue and outputs the final classification result to a result repository. The entire system is deployed using a microservice architecture, allowing each module to be independently extended and updated, ensuring high availability and maintainability.

[0124] In summary, this invention, through the close integration of method and system, constructs a complete multi-dimensional data fusion solution for classifying merchant operation scenarios. This solution not only achieves deep fusion and context awareness of multi-source heterogeneous data technically, but also possesses excellent implementability and scalability in engineering, effectively supporting the demand for high-quality merchant profiles in downstream application scenarios such as financial risk control, precision marketing, and commercial site selection.

[0125] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0126] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for classifying merchant operation scenarios through multi-dimensional data fusion, characterized in that, include: S1: Obtain merchant static attribute data, merchant transaction time series data, merchant text description data, and merchant geospatial data; S2: Modality-specific feature encoding is performed on the four types of data in the following ways to generate static attribute feature vectors, transaction time series feature vectors, semantic text feature vectors, and geospatial feature vectors: merchant static attribute data is input into a multilayer perceptron network, merchant transaction time series data is input into a bidirectional gated recurrent unit network, merchant text description data is input into a pre-trained language model based on a transformer architecture, and merchant geospatial data is input into a convolutional neural network; S3: The cross-modal interactive attention mechanism fuses static attribute feature vectors, transaction time series feature vectors, semantic text feature vectors, and geospatial feature vectors. The cross-modal interactive attention mechanism uses each modal feature vector as a query vector and the remaining modal feature vectors as a set of key value vectors. It calculates attention weights through dot product similarity and normalized exponential function and performs weighted summation to generate four interactively enhanced modal feature vectors in sequence. The vectors are then concatenated and dimensionality reduced to obtain the merchant's multi-dimensional fused feature vector. S4: Construct a merchant relationship graph based on preset geographical proximity rules, customer group similarity rules, and customer sharing rules, and use a graph attention network to enhance the context of the merchant multidimensional fusion feature vector to obtain the final merchant scene representation vector; wherein, the construction of the merchant relationship graph specifically includes: treating each merchant as a node in the graph, and the merchant multidimensional fusion feature vector as the initial attribute of the node; establishing edges between nodes based on preset connection rules, the connection rules including: rule 1, if the geographical straight-line distance between two merchants is less than a preset distance threshold, then establish a geographical proximity edge between their corresponding nodes; rule 2, if the cosine similarity of the target customer profiles of two merchants is greater than a preset similarity threshold, then establish a customer group similarity edge between their corresponding nodes; rule 3, if the customer flow overlap of two merchants within a preset time period is greater than a preset overlap threshold, then establish a customer sharing edge between their corresponding nodes; S5: Input the final merchant scenario representation vector into the preset classifier and output the fine-grained business scenario classification results of the merchant.

2. The merchant operation scenario classification method based on multi-dimensional data fusion according to claim 1, characterized in that, In step S2, modality-specific feature encoding of multi-source heterogeneous merchant data includes: The preprocessed static attribute data of merchants is input into a multilayer perceptron network containing at least two fully connected layers and a nonlinear activation function. Through forward propagation, a static attribute feature vector with fixed dimensions is output. The multidimensional transaction time series is input into a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network captures the long-term dependencies of the transaction series from both forward and reverse time dimensions, and splices the final hidden states of the two directions to form a transaction time series feature vector. The initial word embedding sequence is input into a pre-trained language model based on a transformer architecture. The pre-trained language model captures deep semantic relationships within the text through a multi-head self-attention mechanism and uses the output vector corresponding to the sentence beginning classification label as the semantic text feature vector. The geospatial feature map is input into a convolutional neural network containing multiple convolutional layers, pooling layers, and activation functions. Local spatial patterns with translation invariance are extracted through layer-by-layer convolution operations, and the feature map output from the last convolutional layer is subjected to global average pooling to generate a geospatial feature vector.

3. The merchant operation scenario classification method based on multi-dimensional data fusion according to claim 1, characterized in that, In step S3, generating the merchant's multidimensional fusion feature vector specifically includes: The static attribute feature vector, transaction time sequence feature vector, semantic text feature vector and geospatial feature vector generated in step S2 are used as inputs. One modality feature vector is set as the query vector and the feature vectors of the other three modalities are set as key vectors and value vectors. Calculate the dot product similarity between the query vector and each key vector, and convert the similarity score into attention weights using a normalized exponential function; The obtained attention weights are used to perform a weighted summation on the corresponding set of value vectors to generate a context-aware feature vector that incorporates information from other modalities and is specific to the current query modality. The above calculation process is repeated, using the feature vector of each modality as the query vector, to obtain four modality feature vectors that have been interactively enhanced. The four interactive enhancement modal feature vectors are concatenated in dimension and then reduced in dimension through a linear mapping layer to finally output the merchant's multidimensional fused feature vector.

4. The merchant operation scenario classification method based on multi-dimensional data fusion according to claim 1, characterized in that, The step of inputting the final merchant scenario representation vector into a preset classifier and outputting the merchant's business scenario classification result includes: The final merchant scenario representation vector is input into a classification head consisting of one or more fully connected layers; The last layer of the classification head uses a normalized exponential function as the activation function, and the number of its output nodes is the same as the total number of predefined business scenario categories. The output of the normalized exponential function is a probability distribution vector, where each element represents the probability that the merchant belongs to the corresponding business scenario category; The business scenario category corresponding to the element with the highest probability value is selected as the final classification result output.

5. The merchant operation scenario classification method based on multi-dimensional data fusion according to claim 1, characterized in that, Step S1, which involves obtaining merchant static attribute data, merchant transaction time-series data, merchant text description data, and merchant geospatial data, specifically includes: The merchant static attribute data is obtained from the merchant registration database or a third-party data platform. The merchant static attribute data includes the merchant category code, registered capital, years of establishment, chain brand logo, and approved business scope. By connecting with the payment system interface, all transaction records of the merchant within a preset time window are collected in real time to form the merchant's transaction time sequence data. Each transaction record includes transaction amount, transaction timestamp, payment method and product summary information. By using web crawling technology or application programming interface calls, user reviews, ratings, tags, and business introductions published by merchants themselves are crawled from public comment platforms, social media, and official merchant websites to form the merchant text description data; The precise latitude and longitude coordinates of the merchant are obtained through a geographic information system, and the type, quantity and distribution information of surrounding points of interest are collected within a preset radius centered on the coordinates to form the merchant's geospatial data.

6. The merchant operation scenario classification method based on multi-dimensional data fusion according to claim 2, characterized in that, After acquiring the multi-source heterogeneous merchant data, a data preprocessing step is also included, specifically: Perform one-hot encoding or entity embedding on the static attribute data of the merchants; The merchant transaction time series data is aggregated according to a preset time step, and the total transaction amount, number of transactions, average order value and high-frequency product category distribution within each time step are calculated to form a multi-dimensional transaction time series. The merchant text description data is segmented, stop words are removed, part-of-speech tagging is performed, and stemming is extracted. The text is then converted into an initial word embedding sequence using a word vector model. The merchant geospatial data is processed into a grid, and the density of different types of points of interest within each grid is calculated to form a geospatial feature map.

7. A multi-dimensional data fusion-based merchant operation scenario classification system, applied to the multi-dimensional data fusion-based merchant operation scenario classification method as described in claims 1-6, characterized in that, include: The data acquisition module is used to acquire static attribute data, transaction time series data, text description data, and geospatial data of merchants from merchant registration databases, payment systems, online platforms, and geographic information systems. The feature encoding module integrates a multilayer perceptron network, a bidirectional gated recurrent unit network, a pre-trained language model based on a transformer architecture, and a convolutional neural network, which are used to encode the four modalities of merchant static attribute data, merchant transaction time series data, merchant text description data, and merchant geospatial data into static attribute feature vectors, transaction time series feature vectors, semantic text feature vectors, and geospatial feature vectors, respectively. The interactive fusion module includes a multimodal interactive attention network, which receives the static attribute feature vector, transaction time series feature vector, semantic text feature vector, and geospatial feature vector, and generates a comprehensive merchant multidimensional fusion feature vector through cross-modal query-key-value attention calculation. The context enhancement module is configured to construct a merchant relationship graph based on the geographical proximity, customer group similarity and customer sharing among merchants, and includes a graph attention network for propagating and aggregating the multidimensional fused feature vectors of the merchants on the graph, and outputting a final merchant scene representation vector that incorporates neighborhood context information. The scene classification module includes a fully connected classification head, which receives the final merchant scene representation vector output by the context enhancement module and outputs the final determination result of the business scene category to which the merchant belongs.

Images