Supply delivery overdue risk assessment method, device and equipment based on space-time diagram

Abstract

The invention relates to a supply delivery overdue risk assessment method, device and equipment based on a space-time diagram, and the method comprises the steps: inputting production plant time data and logistics time data into a pre-trained time feature extraction unit for each production plant, obtaining the time splicing feature of each corresponding production plant, and obtaining the time splicing feature of each production plant; a production plant feature representation learning unit is used for extracting dynamic time feature representation of each production plant according to time splicing features and material category feature data by using bidirectional long and short term based on a network in combination with a sequence attention mechanism to construct an undirected weighted graph, node information in the undirected weighted graph is aggregated through the attention network, and the information of each node in the undirected weighted graph is acquired. And finally, an information joint risk assessment unit obtains joint features according to the risk feature representations, the time splicing features and the material category feature data, and a material delivery overdue risk assessment result at the current moment is obtained. By adopting the method, an accurate risk assessment result can be obtained.

Description

Technical Field

[0001] The present application relates to the field of computer information technology, and in particular to a method, device and equipment for assessing the risk of overdue supply delivery based on a spatiotemporal graph. Background Art

[0002] With the booming development of e-commerce goods, the number of manufacturers and transaction volume on these platforms has seen rapid growth. This increase in the number of manufacturers and the diversification of transaction categories has increased the likelihood of claims and disputes arising from delivery and after-sales service between different manufacturers. For e-commerce platforms, effectively managing manufacturer claims risks is crucial, both to protect consumer rights and safeguard the platform's reputation and profitability.

[0003] Traditional risk assessment methods often rely on empirical rules or relatively simple statistical models. For example, manufacturers are roughly graded by observing indicators such as historical claim settlement rates and complaint rates. However, such methods often lack flexibility and accuracy when dealing with complex, large-scale, and dynamically changing data, and are unable to promptly identify potentially high-risk manufacturers. Furthermore, with the improvement of cross-regional logistics networks and the personalization of consumer demand, the risk of material delivery default depends not only on the manufacturer's historical behavior but also on multiple factors such as logistics timeliness, material type characteristics, and product category. Models based solely on a single dimension or static features fail to fully capture potential risks. Summary of the Invention

[0004] Based on this, it is necessary to provide a supply delivery overdue risk assessment method, device and equipment based on a time-space graph that can make accurate estimates based on multi-dimensional information to address the above technical problems.

[0005] A method for assessing supply delivery overdue risk based on a spatiotemporal graph, the method comprising: Obtain relevant data of multiple production plants at the current moment, including production plant time data, logistics time data, and material category feature data; For each production plant, the production plant time data and logistics time data are input into the pre-trained time feature extraction unit, and the corresponding time splicing features of each production plant are extracted through position encoding and self-attention mechanism; Input the time splicing features and material category feature data into the production plant feature representation learning unit, use the bidirectional long-term and short-term network based on the sequential attention mechanism to extract the dynamic time feature representation of each production plant, construct an undirected weighted graph based on the cosine similarity of the dynamic time feature representation of each production plant, and aggregate the information of each node in the undirected weighted graph through the attention network to obtain the risk discrimination representation of each production plant; The risk discrimination representation, time splicing characteristics and material category characteristic data of each production plant are input into the information joint risk assessment unit. The material category feature representation learning model, the logistics information joint model and the production plant information joint model are used to obtain the material category risk feature representation, the logistics risk feature representation and the production plant joint risk feature representation respectively. After splicing the material category risk feature representation, the logistics risk feature representation, the production plant joint risk feature representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

[0006] In one embodiment, the production plant time data includes sales order time series data, the logistics time data includes transportation time series data, and the material category characteristic data includes commodity category and value level.

[0007] In one embodiment, when training the time feature extraction unit, historical production plant time data and historical logistics time data are used as training data of the time feature extraction unit, and the time feature extraction unit is trained by using the time data at a certain moment to predict the sales data of the production plant at the next moment.

[0008] In one embodiment, during training, the temporal feature extraction unit includes a temporal feature encoder, a temporal feature decoder, and an output prediction layer; In the temporal feature encoder, the temporal data is position-encoded respectively, the obtained feature codes are used to obtain corresponding attention values ​​using a multi-head attention mechanism, and feature splicing is performed with the corresponding feature codes to obtain encoder output features; In the temporal feature decoder, position encoding is performed on the encoder output features respectively, and the obtained feature codes are used to obtain corresponding masked attention values ​​using a masked multi-head attention mechanism, and feature splicing is performed with the corresponding feature codes to obtain decoder output features, and the encoder output features and the decoder output features are spliced ​​using a multi-head attention mechanism to obtain spliced ​​features; The splicing features are predicted using the output prediction layer to obtain sales data of the production plant at the next moment, wherein the output prediction layer includes a linear layer and a Softmax layer.

[0009] In one embodiment, only the time feature encoder is retained in the trained time feature extraction unit, and time features are extracted based on the production plant time data and logistics time data.

[0010] In one embodiment, in the production plant characterization learning unit: A bidirectional long-term and short-term network is used to capture the intra-sequence features within each time series in the time splicing features, as well as the inter-sequence features between each time series, to generate hidden states. The hidden states are weighted by the sequence attention mechanism to obtain the dynamic temporal feature representation of each production plant. Based on the dynamic time feature representations, the cosine similarity between the generating plants is calculated, each generating plant is used as a node, and the cosine similarity is used as an edge between corresponding nodes to construct the undirected weighted graph; Using the graph attention network, the feature information of adjacent nodes is aggregated according to each node in the undirected weighted graph network to update the feature representation of each node and obtain the risk discrimination representation of each production plant.

[0011] In one embodiment, in the information joint risk assessment unit: The material category feature representation learning model uses the material category feature data as a query and the risk discrimination representation of the manufacturer as a key / value, and extracts the material category risk feature representation through an attention mechanism; The logistics information joint model extracts logistics dynamic features from logistics time data through a Transformer network and a gated recurrent network, and fuses the logistics dynamic features with material category feature data to obtain the logistics risk feature representation; The production plant information joint model extracts the joint risk feature representation of the production plants through the Transformer network and the gated recurrent network for the time data of the production plants.

[0012] In one embodiment, a supply delivery overdue risk assessment network is constructed based on the pre-trained time feature extraction unit, the production plant feature representation learning unit, and the information joint risk assessment unit; Inputting relevant data of multiple production plants at the current moment into the supply delivery overdue risk assessment network to obtain the material delivery overdue risk assessment result at the current moment; A Dropout strategy is adopted when training the supply delivery overdue risk assessment network.

[0013] The present application also provides a supply delivery overdue risk assessment device based on a time-space graph, the device comprising: The current moment data acquisition module is used to obtain relevant data of multiple production plants at the current moment, including production plant time data, logistics time data, and material category feature data; The time feature extraction module is used to input the production plant time data and logistics time data of each production plant into the pre-trained time feature extraction unit, and extract the corresponding time splicing features of each production plant through position encoding and self-attention mechanism; The risk feature representation extraction module is used to input the time splicing features and material category feature data into the production plant feature representation learning unit, use the bidirectional long-term and short-term network based on the sequential attention mechanism to extract the dynamic time feature representation of each production plant, construct an undirected weighted graph based on the cosine similarity of the dynamic time feature representation of each production plant, and aggregate the information of each node in the undirected weighted graph through the attention network to obtain the risk discrimination representation of each production plant; The risk assessment module is used to input the risk discrimination representation, time splicing characteristics and material category characteristic data of each production plant into the information joint risk assessment unit, and use the material category characteristic representation learning model, the logistics information joint model and the production plant information joint model to obtain the material category risk characteristic representation, the logistics risk characteristic representation and the production plant joint risk characteristic representation respectively. After splicing the material category risk characteristic representation, the logistics risk characteristic representation, the production plant joint risk characteristic representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

[0014] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps in the above-mentioned method for assessing supply delivery overdue risks based on a time-space graph are implemented.

[0015] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for assessing supply delivery overdue risks based on a time-space graph.

[0016] The above-mentioned method, device, and equipment for assessing overdue supply delivery risks based on a spatiotemporal graph inputs the production plant time data and logistics time data for each production plant into a pre-trained time feature extraction unit to obtain the corresponding time splicing features of each production plant. The production plant feature representation learning unit uses the time splicing features and material category feature data to extract the dynamic time feature representation of each production plant using a bidirectional long-short-term network-based, sequential attention mechanism to construct an undirected weighted graph. The attention network aggregates the information of each node in the undirected weighted graph to obtain the risk feature representation of each production plant. Finally, the information joint risk assessment unit uses the risk feature representation, time splicing features, and material category feature data to obtain a joint feature to assess the risk of overdue material delivery at the current moment. This method can obtain accurate risk assessment results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 1 is a flow chart of a method for assessing supply delivery overdue risk based on a time-space graph in one embodiment; Figure 2 Schematic diagram of a framework of a supply delivery overdue risk assessment method based on a spatiotemporal graph in one embodiment; Figure 3 Schematic diagram of a temporal feature extraction unit architecture in one embodiment; Figure 4 1 is a schematic diagram of a data processing flow in a temporal feature encoder according to an embodiment; Figure 5 A schematic diagram of a learning unit structure representing characteristics of a production plant in one embodiment; Figure 6 Schematic diagram of a process for capturing intra-sequence features within each time series in a time splicing feature using a bidirectional long-term and short-term network-based approach in a feature representation learning unit of a production plant in one embodiment; Figure 7 Schematic diagram of the structure of the Transformer model for the joint production plant information in one embodiment; Figure 8 It is a structural block diagram of a supply delivery overdue risk assessment device based on a time-space graph in one embodiment; Figure 9 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0019] In this application, if Figure 1 As shown, a supply delivery overdue risk assessment method based on a spatiotemporal graph is provided, which specifically includes the following steps: Step S100, obtaining relevant data of multiple production plants at the current moment, the relevant data including production plant time data, logistics time data, and material category feature data.

[0020] In step S110, for each production plant, the production plant time data and logistics time data are input into a pre-trained time feature extraction unit, and the corresponding time splicing features of each production plant are extracted through position encoding and self-attention mechanism.

[0021] In step S120, the time splicing features and material category feature data are input into the production plant feature representation learning unit, and the dynamic time feature representation of each production plant is extracted by using a bidirectional long-short-term network combined with a sequential attention mechanism. An undirected weighted graph is constructed based on the cosine similarity of the dynamic time feature representation of each production plant. The information of each node in the undirected weighted graph is aggregated through the attention network to obtain the risk discrimination representation of each production plant.

[0022] In step S130, the risk discrimination representation, time splicing characteristics and material category characteristic data of each production plant are input into the information joint risk assessment unit, and the material category characteristic representation learning model, the logistics information joint model and the production plant information joint model are used to obtain the material category risk characteristic representation, the logistics risk characteristic representation and the production plant joint risk characteristic representation respectively. After splicing the material category risk characteristic representation, the logistics risk characteristic representation, the production plant joint risk characteristic representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

[0023] In this embodiment, a network framework is constructed by using the pre-trained time feature extraction unit, the production plant feature representation learning unit, and the information joint risk assessment unit to mine the spatial correlation information between production plants and the impact of material category factors on material delivery default in the logistics scenario to perform material delivery default risk assessment. The entire network framework is as follows: Figure 2 shown.

[0024] In step S100, the production plant time data includes sales order time series data, the logistics time data includes transportation time series data, and the material category characteristic data includes commodity category, value level and industry data, sales volume, etc.

[0025] In this embodiment, both the production plant time data and the logistics time data are time series data arranged in chronological order.

[0026] In this embodiment, the time feature extraction unit is pre-trained. During the training process, historical production plant time data and historical logistics time data are used as training data for the time feature extraction unit. The time feature extraction unit is trained by using the time data at a certain moment to predict the sales data of the production plant at the next moment.

[0027] like Figure 3As shown, during training, the time feature extraction unit includes a time feature encoder, a time feature decoder and an output prediction layer. In the time feature encoder, the time data is position-encoded respectively, and the obtained feature codes are used to obtain corresponding attention values ​​using a multi-head attention mechanism, and are feature-concatenated with the corresponding feature codes to obtain encoder output features. In the time feature decoder, the encoder output features are position-encoded respectively, and the obtained feature codes are used to obtain corresponding masked attention values ​​using a masked multi-head attention mechanism, and are feature-concatenated with the corresponding feature codes to obtain decoder output features. The encoder output features and the decoder output features are feature-concatenated using a multi-head attention mechanism to obtain concatenated features. Finally, the output prediction layer is used to predict the concatenated features to obtain the sales data of the production plant at the next moment, wherein the output prediction layer includes a linear layer and a Softmax layer.

[0028] In this embodiment, when training the time feature extraction unit, a time data encoder is used to feature encode the production plant's time data and logistics data, and a time data decoder is used to decode these features. Training is performed by predicting the production plant's sales data at the next time point, thereby continuously optimizing the encoding and decoding process.

[0029] Specifically, before uploading the original data of the production plant, the time data and logistics data of the production plant are feature-encoded in the time feature encoder. The encoding process is as follows: Figure 4 At the same time, in the time feature encoder, the historical production plant time data and the historical logistics time data are processed separately, and the processing processes are the same.

[0030] Furthermore, when the time feature encoder performs feature encoding on the time data of the production plant, Time, production plant node Maintains the original time data of the node, including the original factory time data , and the original logistics time data The characteristics of the time data of the production plant are position-coded, and the calculation formula is as follows:

[0031] In the above formula, Indicates the location of the sales order, represents the dimension of the input features, and Indicates parity, Indicates the ordinal number of the dimension. and After position encoding, we get and , defined as follows:

[0032] Encoding characteristics of production plant time data They are used as Query vector, Key vector and Value vector for feature encoding. At time t, they are defined as follows:

[0033] In the above formula, These are the data encoding conversion matrices for Query, Key, and Value. The characteristic encoding of the manufacturer at each moment, its attention value is defined as:

[0034] In the above formula, Indicates the node of the production plant The set of sales orders at the moment. Therefore, The characteristic representation of the production plant node at the time The definitions are as follows:

[0035] In the above formula, The linear transformation matrix encoding the time characteristics of the production plant, As the activation function, a residual connection is performed here to splice the time data encoding features of the production plant , This is the layer normalization operation, thus completing the temporal feature encoding of the production plant.

[0036] Similarly, the logistics time data is encoded into features They are used as Query vector, Key vector and Value vector respectively for feature encoding. Moment, defined as follows:

[0037] In the above formula The transformation matrices encoding Query, Key and Value features respectively.

[0038] for The characteristic encoding of the manufacturer at each moment, its attention value is defined as:

[0039] In the above formula, Indicates the node of the production plant The set of sales orders at the moment. Therefore, The characteristic representation of the production plant node at the time The definitions are as follows:

[0040] In the above formula, is a linear transformation matrix, As the activation function, a residual connection is performed here to splice the time data encoding features of the production plant , This is the layer normalization operation. Thus, the time feature coding of the production plant is completed, and the time feature coding is finally obtained. , defined as follows:

[0041] In the above formula, Represents the temporal features output by the temporal feature encoder.

[0042] Furthermore, a time feature decoder is used to decode the encoder output features, that is, the original factory time data encoder output features , and the encoder output characteristics of the original logistics The two data are encoded by the same position coding as in the above-mentioned time feature encoder to obtain and .

[0043] Next, the location code of the production plant's time data output data The vectors are used as Query vector, Key vector and Value vector for feature encoding. Moment, defined as follows:

[0044] In the above formula, are the encoding matrices of Query, Key and Value respectively. The data output feature encoding of the manufacturer at each moment, its attention value is defined as:

[0045] In the above formula, Indicates the time data of the manufacturer Outputs a collection of sales orders at a time. To add a mask operation. Therefore, The characteristic representation of the production plant node at the time , defined as follows:

[0046] In the above formula, The linear transformation matrix encoding the time characteristics of the production plant, As the activation function, a residual connection is performed here to splice the time data encoding features of the production plant , This is the layer normalization operation, thus completing the temporal feature output encoding of the production plant.

[0047] Similarly, the location encoding of the logistics time data output data The vectors are used as Query vector, Key vector and Value vector for feature encoding. Moment, defined as follows:

[0048] In the above formula, are the feature encoding matrices of Query, Key and Value respectively. The feature encoding of the momentary logistics time data output is defined as:

[0049] In the above formula, Indicates that logistics time data is Outputs a collection of sales orders at a time. To add a mask operation. Therefore, The characteristic representation of the production plant node at the time The definitions are as follows:

[0050] In the above formula, The linear transformation matrix encoding the time characteristics of the production plant, As the activation function, a residual connection is performed here to splice the time data encoding features of the production plant , This is the layer normalization operation. Thus, the logistics time feature output coding is completed. Finally, the production plant time feature output coding and the logistics time feature output coding are spliced ​​to obtain the time feature output coding. , defined as follows:

[0051] Further, in At this moment, the temporal feature encoder outputs the feature , and the temporal feature decoder output features , using an encoder-decoder attention mechanism block, for a certain production plant , the temporal feature decoder output features Set to Query, output features of the temporal feature encoder Set as Key and Value, the specific definitions are as follows:

[0052] In the above formula, The production plants The Query, Key, and Value are transformed into matrices. Then the calculation Time production plant The time feature representation is specifically defined as follows:

[0053] Finally, Send it to the multi-layer perceptron to predict the output prediction of time data at time t , specifically defined as follows:

[0054] In the above formula, is the trainable weight parameter matrix in MLP. Definition The loss function of the pre-training model at this moment is , defined as follows:

[0055] In the above formula, Express The forecast results of the sales data of the manufacturer at that moment, express The actual value of the sales data of the manufacturer at that moment.

[0056] In this embodiment, only the time feature encoder is retained in the trained time feature extraction unit, and time features are extracted based on the production plant time data and logistics time data, thereby improving the effect of feature encoding while protecting the production plant data.

[0057] In step S120, the manufacturer's characteristic representation learning unit structure is as follows Figure 5 As shown in the figure, in this unit, a dynamic feature embedding module and a sequence attention joint module are designed for the production plant node to extract the dynamic features of the production plant time data and capture the key information of claims risk.

[0058] In this embodiment, in the production plant feature representation learning unit, a bidirectional long-term and short-term network is used to capture the intra-sequence features within each time series in the time splicing features, as well as the inter-sequence features between each time series, to generate hidden states, and the hidden states are weighted by the sequence attention mechanism to obtain the dynamic time feature representation of each production plant. The process is as follows: Figure 6 shown.

[0059] Specifically, at time t, the production plant node maintains the time feature encoding after feature encoding , material category characteristic data .

[0060] Furthermore, input the time characteristics of the production plant node Then, in the bidirectional long-term short-term network, the control signal of the forget gate is first generated. , the forget gate is responsible for controlling the previous hidden state The amount of information that should be retained in for use at the next moment can be viewed as follows:

[0061] Specifically, the bidirectional long-term and short-term network can be regarded as a two-layer LSTM neural network, which can be obtained from the bidirectional processing. and , which is defined as follows:

[0062] Finally, and Splice together to get , hidden state is defined as follows:

[0063] The bidirectional long short-term memory network is used to capture the feature coding of the production plant in the time series of the production plant, and the LSTM is used to capture the dynamic features between the time series of the production plant to obtain the temporal dynamic feature coding of the production plant. , specifically defined as follows:

[0064] Furthermore, the output of the bidirectional long short-term memory network, , ,in For the time step of the dynamic feature encoding of the production plant, construct Query, Key, and Value for attention operation, which are specifically defined as follows: . . In the above formula, Represents the time series attention joint transformation matrix of Query and Key. Next, we will calculate The attention vector λ of the feature sequence of the production plant at time t is defined as follows:

[0065] In the calculation of the factory encoded feature vector The material category characteristic data is taken into account , to integrate material category data, which is defined as follows:

[0066] In the above formula, The time step encoding the dynamic characteristics of the production plant.

[0067] Furthermore, based on the dynamic temporal feature representations, the cosine similarity between each production site is calculated. Each site is treated as a node, and the cosine similarity is used as the edge between the corresponding nodes to construct an undirected weighted graph. Using a graph attention network, the feature information of adjacent nodes is aggregated from each node in the undirected weighted graph network to update the feature representation of each node, thereby obtaining the risk feature representation of each production site.

[0068] Specifically, at time t, each production plant is represented as a node in the graph structure, and the cosine similarity distance is calculated using the encoded feature vector from each parent, let and The dynamic time feature representation of two production plants, the similarity calculation between them is defined as follows:

[0069] Then, the threshold is introduced , which is a hyperparameter.

[0070] Specifically, given a simple undirected weighted graph ,in, Represents the collection of production plant nodes, is the set of edges between production plants, where , The feature input of a node is represented by the feature matrix , where N represents the number of features of each production plant, M represents the total number of production plants, and the first Row represents the The feature embedding vector of each factory, the feature input of the edge is defined by the arc table. If and There is a weight between The arc table contains a triple To represent the edge. The dynamic encoder will use the node features of the production plant to initialize the hidden state of the node, ,in Indicates the number of nodes, Indicates the node characteristic number of each node. , the goal is to aggregate information from neighboring production plant nodes and use it to update nodes The similarity coefficient between the central production plant node and the adjacent production plant nodes is calculated. , specifically defined as follows:

[0071] In the above formula, 、 It is a learnable parameter. By linearly transforming each production plant node, it is mapped to a high-dimensional feature space to improve its expressive power. Represents the transformed production plant node set, Represents the connection operation, which splices the transformed production plant features into ,Then calculate the attention coefficient between adjacent production plant nodes, It is a single-layer feedforward neural network, represented by a weight vector: , which maps the concatenated production plant features to a real number as the attention coefficient. Function normalizes the attention coefficient and uses activation function, which, preferably, has a negative input slope =0.35), specifically defined as follows:

[0072] In the above formula, Node for the production plant Next, the risk discrimination representation of the central production plant node is updated based on the neighboring nodes of the central production plant. , using a multi-head attention mechanism, specifically defined as follows:

[0073] In the above formula, is the activation function, Node for adjacent production plants No. The feature correlation transformation matrix of each attention head.

[0074] In step S130, based on the full-process information of commodity purchase, in the information joint risk assessment module, the present invention designs a material category feature representation learning model, a logistics information joint Transformer model, and a production plant information joint Transformer model to supplement the risk feature representation of production plant nodes, making it more in line with the actual characteristics of material delivery default risk assessment in logistics scenarios.

[0075] In this embodiment, the information-joint risk assessment unit uses the material category feature representation learning model, using the material category feature data as a query and the production site risk discrimination representation as a key / value pair, to extract the material category risk feature representation through an attention mechanism. The logistics information joint model uses a Transformer network and a gated recurrent network to extract dynamic logistics features from logistics time data, and then fuses these dynamic logistics features with the material category feature data to obtain a logistics risk feature representation. The production site information joint model uses a Transformer network and a gated recurrent network to extract a joint risk feature representation for the production site from the production site time data.

[0076] Specifically, the material category feature representation learning model is The risk identification of each production plant node can be obtained at any time Its core data is usually regarded as Query, while other data are called Key and Value respectively. At time t, the material category characteristics are expressed as As a query, the manufacturer's coded features As Key and Value, they are defined as follows:

[0077] In the above formula, Represents the material category encoding conversion matrix of Query, Key and Value respectively. Next, the feature representation of the material category feature at time t is calculated and recorded as , the specific calculation is as follows:

[0078] Therefore, we obtain the learning model by characterizing the material category features. Material category characteristics at the moment .

[0079] Specifically, the logistics information is combined with the Transformer model to encode the logistics features at time t use The model embeds dynamic features within and between logistics time series, capturing the dynamic feature encoding of logistics. , the process is as follows Figure 6 The specific definitions are as follows:

[0080] Extract important information from the logistics dynamic feature coding of all logistics time steps, and then obtain the logistics dynamic feature coding. Afterwards, the sequence attention mechanism is used to obtain the logistics encoding feature vector .

[0081] Representing material category characteristics and logistics encoding feature vector Combined with gated neural networks Capturing cyclical patterns in logistics information.

[0082] At time t, the updated gate vector of the logistics feature is first generated , which controls the hidden state at the previous moment How much information is retained and used at the next moment is defined as follows:

[0083] In the above formula, To update the gate weight matrix, , is the bias term, The role of the activation function is to limit the output to Within this specific range, Represents the feature vector concatenation. Subsequently, the reset gate vector is generated , this vector controls how much historical information is retained and determines what needs to be forgotten. It is defined as follows:

[0084] In the above formula, To reset the gate weight matrix, is the bias term. In order to fuse the moments Material category characteristics , logistics coding feature vector And the hidden state at the last moment , constructing candidate hidden states The GRU process can be simplified into the following form:

[0085] like Figure 7 As shown in the figure, the structure of the Transformer model for the production plant information is used to encode the logistics characteristics at time t. Utilized The module embeds dynamic features within and between production plant time series, capturing the dynamic feature coding of production plants , which is defined as follows:

[0086] By obtaining the dynamic feature code of the manufacturer , further using the sequence attention mechanism to obtain the feature vector encoded by the manufacturer Next, we use the gated neural network to capture the periodic model of the production plant information and obtain the risk characteristic representation of the production plant

[0087] Obtain the logistics risk characteristic representation at time t and risk characteristics of manufacturers At the same time, the risk discrimination representation of the production plant output from the production plant representation learning model is obtained , and then these feature representations are concatenated and passed through a multi-layer perceptron ( ) to process, thereby achieving The risk assessment and prediction of material delivery default at the time is defined as follows:

[0088] In the above formula, are the learnable weight parameters of the multi-layer perceptron (MLP) model in information joint risk assessment.

[0089] In this embodiment, a supply delivery overdue risk assessment network can also be constructed based on the pre-trained time feature extraction unit, production plant feature representation learning unit, and information joint risk assessment unit. In this way, the relevant data of multiple production plants at the current moment can be directly input into the supply delivery overdue risk assessment network to obtain the current material delivery overdue risk assessment result. When training the supply delivery overdue risk assessment network, the following loss function is used:

[0090] In the above formula, in order to prevent overfitting, the Dropout strategy is also adopted during the training process.

[0091] In this paper, the proposed method is also implemented on an online commodity trading platform. Changes in the business environment, such as large-scale promotions, holidays, and special events, often lead to significant fluctuations in sales patterns. These changes are usually accompanied by abnormal increases or decreases in sales volume in the short term, and their nature and magnitude are highly time-sensitive and uncertain. Due to the complexity of these fluctuations, traditional sales forecasting models often have difficulty effectively capturing such abnormal patterns, especially when faced with special periods that deviate from the normal pattern, the model's predictive ability may be significantly reduced. Most traditional forecasting models are usually trained based on historical sales data and assume that sales data exhibits a certain degree of regularity and stability, ignoring the sudden fluctuations that may be caused by factors such as promotions or holidays.

[0092] Specifically, factors such as promotions, holidays, and special events directly influence consumer purchasing decisions, causing sales data to exhibit characteristics significantly different from normal patterns in the short term. Therefore, effectively integrating these external anomalies into forecasting models becomes a key challenge in improving sales forecast accuracy and model generalization. To address this issue, forecasting models must not only identify long-term trends but also possess sufficient dynamic adaptability to cope with the frequent and unpredictable changes in the business environment.

[0093] Therefore, during a shopping event, we conducted experiments using this method to evaluate the model's performance and effectiveness in a large-scale promotion. As shown in Table 1, the experimental results demonstrate that during the event, our proposed method, compared to the HST-GT method, maintained relatively stable performance on the MAE, MSE, MAPE, and SMAPE evaluation metrics, with minimal fluctuations, demonstrating strong adaptability and reliability.

[0094] Table 1 Experimental results of this method and traditional methods

[0095] The aforementioned spatiotemporal graph-based supply delivery overdue risk assessment method addresses the complexities of claims assessment and designs Transformer-based pre-training modules for the spatiotemporal graph: a temporal feature extraction module, a production site node representation learning module, and an information joint risk assessment module. While protecting production site data security, the pre-training module effectively ensures the accuracy of claims risk assessment by learning data features from e-commerce and logistics platforms. Subsequently, the production site node representation learning module learns representations for each production site node based on a production site feature representation learning model and a production site risk discrimination model. These representations are then fed into the information joint risk assessment module. This module, tailored to logistics scenarios, incorporates a material category feature representation learning model, a logistics information joint Transformer model, and a production site information joint Transformer model. This module leverages full-process logistics data and material category factors to mine spatial correlations between production sites, thereby achieving more accurate material delivery default risk assessment. This method enables production site-specific claims risk assessment and prediction and is applicable to a variety of e-commerce and logistics scenarios, including online retail platforms, cross-border e-commerce, and O2O delivery services. While safeguarding manufacturers' privacy and data security, this method integrates dynamic data on material categories, logistics, and manufacturers' own processes to comprehensively assess multiple key factors that may lead to claims disputes. This helps platforms and manufacturers promptly identify potential risks and accurately estimate claims costs. Compared to existing technologies, this method offers a significant advantage in balancing data security and algorithmic interpretability, significantly improving risk monitoring efficiency, optimizing resource allocation, and enhancing user experience. Furthermore, it can provide regulators with more intuitive data support, promoting the healthy and sustainable development of the e-commerce ecosystem.

[0096] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0097] In one embodiment, Figure 8As shown, a supply delivery overdue risk assessment device based on a spatiotemporal graph is provided, comprising: a current moment data acquisition module 200, a time feature extraction module 210, a risk feature representation extraction module 220, and a risk assessment module 230, wherein: The current moment data acquisition module 200 is used to obtain relevant data of multiple production plants at the current moment, including production plant time data, logistics time data, and material category feature data; The time feature extraction module 210 is used to input the production plant time data and logistics time data of each production plant into the pre-trained time feature extraction unit, and extract the corresponding time splicing features of each production plant through position encoding and self-attention mechanism; The risk feature representation extraction module 220 is used to input the time splicing features and material category feature data into the production plant feature representation learning unit, use a bidirectional long-term and short-term network based on the sequential attention mechanism to extract the dynamic time feature representation of each production plant, construct an undirected weighted graph based on the cosine similarity of the dynamic time feature representation of each production plant, and aggregate the information of each node in the undirected weighted graph through the attention network to obtain the risk discrimination representation of each production plant; The risk assessment module 230 is used to input the risk discrimination representation, time splicing characteristics and material category characteristic data of each production plant into the information joint risk assessment unit, and use the material category characteristic representation learning model, the logistics information joint model and the production plant information joint model to obtain the material category risk characteristic representation, the logistics risk characteristic representation and the production plant joint risk characteristic representation respectively. After splicing the material category risk characteristic representation, the logistics risk characteristic representation, the production plant joint risk characteristic representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

[0098] Regarding the specific limitations of the supply delivery overdue risk assessment device based on the time-space graph, please refer to the limitations of the supply delivery overdue risk assessment method based on the time-space graph above, which will not be repeated here. The various modules in the above-mentioned supply delivery overdue risk assessment device based on the time-space graph can be implemented in whole or in part through software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0099] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure may be shown in Figure 9. The computer device includes a processor, memory, a network interface, a display screen, and an input device connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device communicates with an external terminal via a network connection. When executed by the processor, the computer program implements a method for assessing supply delivery overdue risks based on a spatiotemporal graph. The display screen of the computer device may be a liquid crystal display or an electronic ink display. The input device of the computer device may be a touchscreen covering the display screen, keys, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse.

[0100] Those skilled in the art will understand that Figure 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0101] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented: Obtain relevant data of multiple production plants at the current moment, including production plant time data, logistics time data, and material category feature data; For each production plant, the production plant time data and logistics time data are input into the pre-trained time feature extraction unit, and the corresponding time splicing features of each production plant are extracted through position encoding and self-attention mechanism; Input the time splicing features and material category feature data into the production plant feature representation learning unit, use the bidirectional long-term and short-term network based on the sequential attention mechanism to extract the dynamic time feature representation of each production plant, construct an undirected weighted graph based on the cosine similarity of the dynamic time feature representation of each production plant, and aggregate the information of each node in the undirected weighted graph through the attention network to obtain the risk discrimination representation of each production plant; The risk discrimination representation, time splicing characteristics and material category characteristic data of each production plant are input into the information joint risk assessment unit. The material category feature representation learning model, the logistics information joint model and the production plant information joint model are used to obtain the material category risk feature representation, the logistics risk feature representation and the production plant joint risk feature representation respectively. After splicing the material category risk feature representation, the logistics risk feature representation, the production plant joint risk feature representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

[0102] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: Obtain relevant data of multiple production plants at the current moment, including production plant time data, logistics time data, and material category feature data; For each production plant, the production plant time data and logistics time data are input into the pre-trained time feature extraction unit, and the corresponding time splicing features of each production plant are extracted through position encoding and self-attention mechanism; Input the time splicing features and material category feature data into the production plant feature representation learning unit, use the bidirectional long-term and short-term network based on the sequential attention mechanism to extract the dynamic time feature representation of each production plant, construct an undirected weighted graph based on the cosine similarity of the dynamic time feature representation of each production plant, and aggregate the information of each node in the undirected weighted graph through the attention network to obtain the risk discrimination representation of each production plant; The risk discrimination representation, time splicing characteristics and material category characteristic data of each production plant are input into the information joint risk assessment unit. The material category feature representation learning model, the logistics information joint model and the production plant information joint model are used to obtain the material category risk feature representation, the logistics risk feature representation and the production plant joint risk feature representation respectively. After splicing the material category risk feature representation, the logistics risk feature representation, the production plant joint risk feature representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

[0103] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A supply delivery overdue risk assessment method based on a spatiotemporal graph, characterized in that: The method comprises: Obtain relevant data of multiple production plants at the current moment, including production plant time data, logistics time data, and material category feature data; For each production plant, the production plant time data and logistics time data are input into the pre-trained time feature extraction unit, and the corresponding time splicing features of each production plant are extracted through position encoding and self-attention mechanism; Input the time splicing features and material category feature data into the production plant feature representation learning unit, use the bidirectional long-term and short-term network based on the sequential attention mechanism to extract the dynamic time feature representation of each production plant, construct an undirected weighted graph based on the cosine similarity of the dynamic time feature representation of each production plant, and aggregate the information of each node in the undirected weighted graph through the attention network to obtain the risk discrimination representation of each production plant; The risk discrimination representation, time splicing characteristics and material category characteristic data of each production plant are input into the information joint risk assessment unit. The material category feature representation learning model, the logistics information joint model and the production plant information joint model are used to obtain the material category risk feature representation, the logistics risk feature representation and the production plant joint risk feature representation respectively. After splicing the material category risk feature representation, the logistics risk feature representation, the production plant joint risk feature representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

2. The supply delivery overdue risk assessment method based on a spatiotemporal graph according to claim 1 is characterized in that: The production plant time data includes sales order time series data, the logistics time data includes transportation time series data, and the material category characteristic data includes commodity category and value level.

3. The supply delivery overdue risk assessment method based on a spatiotemporal graph according to claim 2 is characterized in that: When training the time feature extraction unit, historical production plant time data and historical logistics time data are used as training data for the time feature extraction unit. The time feature extraction unit is trained by using the time data at a certain moment to predict the sales data of the production plant at the next moment.

4. The method for assessing supply delivery overdue risk based on a spatiotemporal graph according to claim 3 is characterized in that: During training, the temporal feature extraction unit includes a temporal feature encoder, a temporal feature decoder, and an output prediction layer; In the temporal feature encoder, the temporal data is position-encoded respectively, the obtained feature codes are used to obtain corresponding attention values ​​using a multi-head attention mechanism, and feature splicing is performed with the corresponding feature codes to obtain encoder output features; In the temporal feature decoder, position encoding is performed on the encoder output features respectively, and the obtained feature codes are used to obtain corresponding masked attention values ​​using a masked multi-head attention mechanism, and feature splicing is performed with the corresponding feature codes to obtain decoder output features, and the encoder output features and the decoder output features are spliced ​​using a multi-head attention mechanism to obtain spliced ​​features; The splicing features are predicted using the output prediction layer to obtain sales data of the production plant at the next moment, wherein the output prediction layer includes a linear layer and a Softmax layer.

5. The supply delivery overdue risk assessment method based on a spatiotemporal graph according to claim 4 is characterized in that: The trained time feature extraction unit only retains the time feature encoder, and extracts time features based on the production plant time data and logistics time data.

6. The supply delivery overdue risk assessment method based on a spatiotemporal graph according to claim 5 is characterized in that: In the feature representation learning unit of the production plant: A bidirectional long-term and short-term network is used to capture the intra-sequence features within each time series in the time splicing features, as well as the inter-sequence features between each time series, to generate hidden states. The hidden states are weighted by the sequence attention mechanism to obtain the dynamic temporal feature representation of each production plant. Based on the dynamic time feature representations, the cosine similarity between the generating plants is calculated, each generating plant is used as a node, and the cosine similarity is used as an edge between corresponding nodes to construct the undirected weighted graph; Using the graph attention network, the feature information of adjacent nodes is aggregated according to each node in the undirected weighted graph network to update the feature representation of each node and obtain the risk discrimination representation of each production plant.

7. The supply delivery overdue risk assessment method based on a spatiotemporal graph according to claim 6 is characterized in that: In the information joint risk assessment unit: The material category feature representation learning model uses the material category feature data as a query and the risk discrimination representation of the manufacturer as a key / value, and extracts the material category risk feature representation through an attention mechanism; The logistics information joint model extracts logistics dynamic features from logistics time data through a Transformer network and a gated recurrent network, and fuses the logistics dynamic features with material category feature data to obtain the logistics risk feature representation; The production plant information joint model extracts the joint risk feature representation of the production plants through the Transformer network and the gated recurrent network for the time data of the production plants.

8. The method for assessing supply delivery overdue risk based on a spatiotemporal graph according to any one of claims 1 to 7, characterized in that: A supply delivery overdue risk assessment network is constructed based on the pre-trained time feature extraction unit, production plant feature representation learning unit, and information joint risk assessment unit. Inputting relevant data of multiple production plants at the current moment into the supply delivery overdue risk assessment network to obtain the material delivery overdue risk assessment result at the current moment; A Dropout strategy is adopted when training the supply delivery overdue risk assessment network.

9. A supply delivery overdue risk assessment device based on a time-space graph, characterized in that: The device comprises: The current moment data acquisition module is used to obtain relevant data of multiple production plants at the current moment, including production plant time data, logistics time data, and material category feature data; The time feature extraction module is used to input the production plant time data and logistics time data of each production plant into the pre-trained time feature extraction unit, and extract the corresponding time splicing features of each production plant through position encoding and self-attention mechanism; The risk feature representation extraction module is used to input the time splicing features and material category feature data into the production plant feature representation learning unit, use the bidirectional long-term and short-term network based on the sequential attention mechanism to extract the dynamic time feature representation of each production plant, construct an undirected weighted graph based on the cosine similarity of the dynamic time feature representation of each production plant, and aggregate the information of each node in the undirected weighted graph through the attention network to obtain the risk discrimination representation of each production plant; The risk assessment module is used to input the risk characteristic representation, time splicing characteristics and material category characteristic data of each production plant into the information joint risk assessment unit, and use the material category characteristic representation learning model, the logistics information joint model and the production plant information joint model to obtain the material category risk characteristic representation, the logistics risk characteristic representation and the production plant joint risk characteristic representation respectively. After splicing the material category risk characteristic representation, the logistics risk characteristic representation, the production plant joint risk characteristic representation and the risk discrimination representation, they are processed through a multi-layer perceptron to obtain the material delivery overdue risk assessment result at the current moment.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images