Multi-category material demand prediction system under emergent public health event

Abstract

The invention discloses a multi-category material demand prediction system under an emergent public health event, and the system at least comprises a time sequence demand evolution coding module, a double-correlation learning module, and a regional joint prediction module. The time sequence demand evolution coding module is used for capturing a multi-time-scale demand mode from historical demand records and space-time emergent public health event characteristics; the double-association learning module is used for respectively learning experience association and adaptive association between categories; and the regional joint prediction module obtains fine-grained future demands by aggregating the demand mode output by the time sequence demand evolution coding module and the association mode output by the double-association learning module based on the graph diffusion network. According to the method, the regional characteristics and the dynamic characteristics of health event evolution are comprehensively considered, so that the daily demand quantity of each regional material category is accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer intelligent computing and application, and in particular, relates to a multi-category material demand prediction technology under public health emergencies based on a multi-scale relationship learning framework, and specifically relates to a multi-category material demand prediction system under public health emergencies. Background Art

[0002] In the past few years, public health emergencies (PHEs) have had a significant impact on people's daily lives and demand characteristics, which in turn has brought new challenges to the online and offline demand forecasting of different categories of products. For example, in the early stages of PHEs, online shopping platforms faced an explosive demand for medical supplies, and the shortage of supplies and increased pressure on logistics supply seriously affected social order. To avoid the recurrence of such situations, accurately predicting the multi-category demand under PHEs is an important prerequisite for ensuring material supply, making decisions on product pre-stocking, and allocating regional transportation capacity.

[0003] User needs are a comprehensive reflection of multiple intuitive and complex needs, which depend on user psychology, social development and personal preferences. In other words, in addition to users' personal preferences, social events will also have an impact on their needs. The evolution of needs caused by PHEs is mainly reflected in the following three points:

[0004] First, user preferences change. Usually, user demand tends to have seasonal regularity; in public health emergencies, users' daily needs or panic demands will show preferences for necessities. Some works have proposed forecasting solutions under normal circumstances, but they fail to capture the shift in demand patterns under emergency situations.

[0005] Second, changes in category associations. As the epidemic spreads, there are obvious fluctuations in the complementary functions (health products and baby products) or substitution functions (fresh food and food) between different categories of demand, such as Figure 1 As shown, Figure 1 It shows the changes in correlation during the outbreak of public health emergencies. Most recent studies predict single-category demand without considering the correlation between multiple categories.

[0006] Third, the regional demand distribution has shifted. Normally, online demand is mainly distributed throughout residential and office areas. However, during public health emergencies, due to the decrease in population in urban areas, most of the demand is biased towards residential areas. We can see from Figure 2 Clear differences in the spatial distribution of regional demand can be seen.

[0007] Therefore, accurate prediction of multi-category demand requires not only modeling of users’ dynamic preferences, but also modeling of complex spatiotemporal event evolution. In addition, user purchase records from online shopping platforms can also be used to clearly construct associations between categories. Summary of the invention

[0008] The present invention is aimed at the problems of unclear emergency scenario modeling and unclear category association in the prior art, and provides a multi-category material demand forecasting system under public health emergencies, which at least includes a temporal demand evolution coding module, a dual association learning module and a regional joint forecasting module. The temporal demand evolution coding module is used to capture multi-time scale demand patterns from historical demand records and spatiotemporal public health emergency characteristics; the dual association learning module is used to learn the empirical association and adaptive association between categories respectively; the regional joint forecasting module is based on a graph diffusion network, and obtains fine-grained future demand by aggregating the demand pattern output by the temporal demand evolution coding module and the association pattern output by the dual association learning module. The present invention comprehensively considers the characteristics of the region itself and the dynamic characteristics of the evolution of health events to accurately predict the daily demand for material categories in each region.

[0009] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a multi-category material demand forecasting system under public health emergencies, comprising at least a temporal demand evolution coding module, a dual association learning module and a regional joint forecasting module,

[0010] The temporal demand evolution encoding module is used to capture multi-time scale demand patterns from historical demand records and spatiotemporal characteristics of public health emergencies;

[0011] The dual association learning module is used to learn the empirical association and the adaptive association between categories respectively; the empirical association includes a long-term empirical association matrix and a short-term empirical association matrix; the adaptive association includes an adaptive relationship matrix, which is learned by end-to-end stochastic gradient descent;

[0012] The regional joint prediction module: based on the graph diffusion network, obtains fine-grained future demand by aggregating the demand pattern output by the temporal demand evolution encoding module and the association pattern output by the dual association learning module.

[0013] As an improvement of the present invention, the temporal demand evolution coding module at least includes a multi-time scale encoder and an event-related demand evolution encoder.

[0014] The multi-time scale encoder includes a short-term demand encoder and a long-term demand encoder, wherein:

[0015] In the short-term demand encoder, LSTM is used to capture the short-term demand pattern in the historical records to obtain the short-term demand features. A self-attention mechanism is introduced to weight the temporal features of each region, and the global features of the region are obtained by summing all the weighted local features of the region.

[0016] In the long-term demand encoder, global features of the entire time series are extracted through convolution;

[0017] In the event-related demand evolution encoder, LSTM is used to extract features, and the neural Hawkes process is introduced to simulate the event impact intensity that changes over time to perform demand pattern fusion.

[0018] As another improvement of the present invention, the feature extraction in the long-term demand encoder is specifically as follows: input the long-term demand sequence but

[0019]

[0020] in, Represents sequence characteristics, represents the required features, sigmoid(·) and tanh(·) are activation functions, Conv(·) is the temporal convolution operation, and the convolution kernel is k; w s 、w l , b are trainable parameters.

[0021] As another improvement of the present invention, the output of the timing requirement evolution encoding module is:

[0022]

[0023] in, Indicates different types of demand characteristics, represents the demand evolution characteristics, and the weight parameter β of each scale encoder represents the weight parameter of each scale encoder. Represents the weight parameter of the feature, s represents short-term features, l represents long-term features, and e represents event evolution features.

[0024] As another improvement of the present invention, the adaptive matrix is ​​calculated as follows:

[0025]

[0026] Among them, the softmax(·) function is used as an activation function to eliminate weak connections between categories; the ReLU(·) function is used as an activation function to normalize the adaptive adjacency matrix; E1 and E2 are vector representations of randomly initialized category pairs.

[0027] Compared with the prior art, the technical advantages and effects of the present invention are as follows: This method accurately predicts the multi-category demand under public health emergencies, which is an important prerequisite for ensuring material supply, making decisions on product pre-stocking, allocating regional transportation capacity, etc. Its technical advantages are: using high-precision prediction models and big data analysis technology to achieve high-precision prediction of different categories of demand; integrating real-time data monitoring systems to respond quickly to emergencies and adjust supply chain strategies in a timely manner; and conducting multi-dimensional analysis through multi-source data fusion to improve the comprehensiveness and accuracy of predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic diagram of the association between categories during the outbreak of public health emergencies;

[0029] Figure 2 This is a schematic diagram of the spatial distribution of demand during the outbreak of public health emergencies;

[0030] Figure 3 This is a schematic diagram of a multi-category demand joint prediction framework during a public health emergency of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0032] Example 1

[0033] A multi-category material demand forecasting system for public health emergencies, such as Figure 3 As shown, it at least includes a time series demand evolution encoding module, a dual association learning module and a regional joint prediction module; the time series demand evolution encoding module is used to capture multi-time scale demand patterns from historical demand records and spatiotemporal characteristics of public health emergencies; the dual association learning module is used to learn empirical associations and adaptive associations between categories respectively; the regional joint prediction module, from the perspective of geographic space, obtains fine-grained future demand by aggregating previous demand patterns and association patterns in each region.

[0034] Among these three important modules, the temporal demand evolution coding module: considering that the user's demand pattern is affected by PHEs, and this impact will be aggravated or weakened, the Hawkes process is used to model the user's demand pattern under PHEs from three aspects: short-term demand, long-term demand and basic demand.

[0035] Dual association learning module: In order to use the category relationship information at different time scales, the integrated knowledge is used to enhance the inter-category relationship in multi-category demand forecasting.

[0036] Regional Joint Prediction Module: In order to capture highly salient patterns and features in the spatial domain, a diffuse convolutional network is used to extend the prediction task to joint prediction between regions.

[0037] Timing requirement evolution coding module:

[0038] Temporal patterns vary greatly at different time window scales, for example, typical user preferences will reflect long-term demand rather than specific categories of products, while recent records show short-term demand. In the case of PHE, there is also an important factor that affects the immediate demand for certain related products, such as healthcare and pharmaceutical products. However, the evolution of the event itself will also change the demand characteristics. Therefore, we need to capture the dynamic evolution of all categories of demand to understand how the state of PHE affects preference shifts. Therefore, there are two components in the demand evolution encoding module: Multi-time scale encoder Event-related demand evolution encoder ENC ε .

[0039] In order to model the multi-granularity patterns in historical demand records, two independent time series encoders are designed, namely the short-term demand encoder and the long-term demand encoder. The time series is segmented into short-term demand series and long-term demand series by time windows of different sizes. n} is recorded as the demand sequence of n regions, where for region r i , d i ={d i,1 ,d i,2 ,…,d i,m} represents the demand sequence of m categories in the region. The time series is divided into short-term demand sequences by time windows of different sizes. and long-term demand sequence

[0040] Short-term demand encoder: Use LSTM to capture short-term demand patterns in historical records. For region r i ,we will Input LSTM to obtain short-term demand features, that is, Where m is the number of categories.

[0041] Considering the impact of inter-regional demands on local needs, a self-attention mechanism is introduced to weight the temporal features of each region, and the region r is obtained by summing up all weighted regional local features. i The global features of the region r i and region r j The self-attention weight for category u is defined as

[0042]

[0043] Among them, w q and w k is a trainable parameter. Calculate the weighted sum to get the region r i Global features:

[0044]

[0045] Where ⊙ represents the element-wise multiplication operation.

[0046] The long-term demand encoder and the event-related demand evolution encoder use the same attention mechanism to calculate the region r i Global features and

[0047] Long-term demand encoder: Based on the long-term historical demand records of users, explore the long-term demand or seasonal demand of various categories. Different from short-term demand, TCN is used to extract time features on the time axis. Compared with LSTM, this model extracts global features on the entire time series through convolution, while not forgetting historical information, and is more effective in processing long-term time series data. Input long-term demand sequence

[0048]

[0049] Among them, sigmoid(·) and tanh(·) are activation functions, Conv(·) is the temporal convolution operation, and the convolution kernel is k; w q 、w k , b are trainable parameters.

[0050] In the event-related demand evolution encoder, use Represents a set of event-related requirements, where c i Indicates the characteristics of public health emergencies, s i ={s i,1 ,s i,2 ,…,s i,m} represents region r i The demand sequence of m categories in the PHEs. To model the evolution of PHEs, the following statistics are selected as regional features: the number of confirmed cases of public health emergencies per day. The sequence length k is related to the period of PHEs and is set to 14 in this method.

[0051] Similar to the short-term demand encoder, LSTM is also used to extract features. PHEs feature c i and the vector representation of the demand characteristics s i They are calculated as:

[0052]

[0053] The resulting vector is represented as and Stitching:

[0054]

[0055] Since the influence of PHEs decreases over time and the decay rates are different at different time scales, a neural Hawkes process is introduced to simulate the time-varying intensity of event impacts. s (t), λ l (t), λ e (t) represents the weight of different demand encoders. s (t) is expressed as:

[0056]

[0057] Among them, t0 is the occurrence time of PHEs, t is the current time; w s is a trainable parameter; ∈ is an excitation factor; γ is a decay rate. The output of the temporal demand evolution encoding module is:

[0058]

[0059] Among them, β is the weight parameter of each scale encoder.

[0060] Dual-associative learning modules:

[0061] Dynamic category demand usually depends on the evolution of PHEs. The present invention includes a dual-association learning module to analyze the correlation between categories from two aspects. The dual-relationship learning module analyzes and mines the correlation between categories from two aspects: empirical relationship and adaptive relationship in order to utilize category association information at different time scales and improve the prediction accuracy of category demand.

[0062] Experience relationship matrix: Experience relationship matrix A ε It aims to construct the general category relationship under PHEs. Since the demand pattern capture module consists of multi-scale long-short-term demand encoders, A ε Can also be divided into and Represent short-term and long-term correlations respectively. Based on the purchase records of the e-commerce platform, the relationship strength between categories is extracted with 7 days as the segmentation to obtain the short-term empirical correlation matrix We then counted the number of people to analyze the co-occurrence of products, which revealed the complementary relationships between the categorized needs. Over a longer period of time – 90 days – calculate the co-occurrence frequency between category pairs.

[0063] Adaptive Relationship Matrix: Adaptive Correlation Matrix The implicit correlation between multivariate time series is explored by updating the association matrix, which does not require additional prior knowledge and is learned through end-to-end stochastic gradient descent. The vector representations E1 and E2 of the randomly initialized category pairs are iteratively updated during model training. E1 represents the source category node and E2 represents the target category node. The adaptive matrix is ​​calculated as follows:

[0064]

[0065] Among them, the softmax(·) function is used as an activation function to eliminate weak connections between categories; the ReLU(·) function is used as an activation function to normalize the adaptive adjacency matrix. Similar to the empirical association matrix, the adaptive association matrix is ​​also divided into a short-term adaptive association matrix and a long-term adaptive association matrix. The elements in the matrix reflect the correlation and association strength between categories. Using the matrix, a graph-based relationship fusion method is further established, and the association matrix is ​​mapped to the relationship graph.

[0066] Relationship modeling: The elements in the matrix reflect the correlation or association strength between categories. A graph-based relationship fusion method is further established, and the association matrix is ​​mapped into a relationship graph. ε and As the original incidence matrix, the relationship graph G is constructed ε and The nodes represent categories, and the edges represent the strength of the correlation between two category nodes.

[0067] For node v i , its adjacent node set and edge set are expressed as and φ(v,e) is the node relationship fusion function, where v is the node vector and e is the edge vector. i The characteristic expression is:

[0068]

[0069] in, Aggregation is a way to fuse information from neighboring nodes with information from a central node.

[0070] Regional joint prediction module:

[0071] Adjacent urban areas often show similar event evolution patterns, especially in spatiotemporal public health emergencies, which follow certain propagation laws. Therefore, it becomes crucial to capture the dependencies between regions when characterizing the evolving impact of public health emergencies. We design a regional joint prediction module based on a graph diffusion network to further learn the joint features of each region. We construct a regional graph And the adjacency matrix weighted by the vertex distance function is expressed as The diffusion convolution operation is denoted as Conv d , to perform the diffusion process between nodes and generate new feature representations:

[0072]

[0073] Among them, K represents the receptive field width of the filter, the number of units corresponds to the number of filters, and θ k is the parameter set of the filter, i represents the region ID, represents the hidden variable of region i, D -1 A S Represents the transfer matrix of the graph diffusion process. Finally, the activation function is used to represent the hidden layer Normalize:

[0074]

[0075] In addition, we use a fully connected network as the output layer for multi-class demand prediction and simultaneously calculate the final prediction value of multi-class demand for all regions. The prediction value is We use mean square error (MSE) as the prediction loss function, expressed as in Y is the actual demand value, y i,j ∈Y.

[0076] In summary, in the system of the present invention, the temporal demand evolution coding module proposes a novel demand evolution coding method, which performs multi-scale encoding on historical sequences to effectively capture the immediate demand and long / short-term demand evolution caused by public health events from historical records; the dual association learning module proposes a new type of association learning method to enhance the understanding of the association between categories. The system of the present invention takes into account the characteristics of the region itself and the dynamic characteristics of the evolution of health events to accurately predict the daily demand for material categories in each region.

[0077] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A multi-category material demand forecasting system under public health emergencies, characterized by: At least including the temporal demand evolution encoding module, the dual association learning module and the regional joint prediction module, The temporal demand evolution encoding module is used to capture multi-time scale demand patterns from historical demand records and spatiotemporal characteristics of public health emergencies; The dual association learning module is used to learn the empirical association and the adaptive association between categories respectively; the empirical association includes a long-term empirical association matrix and a short-term empirical association matrix; the adaptive association includes an adaptive relationship matrix, which is learned by end-to-end stochastic gradient descent; The regional joint prediction module: based on the graph diffusion network, obtains fine-grained future demand by aggregating the demand pattern output by the temporal demand evolution encoding module and the association pattern output by the dual association learning module.

2. The system for forecasting the demand for multiple categories of materials under public health emergencies according to claim 1, characterized in that: The temporal demand evolution coding module at least includes a multi-time scale encoder and an event-related demand evolution encoder. The multi-time scale encoder includes a short-term demand encoder and a long-term demand encoder, wherein: In the short-term demand encoder, LSTM is used to capture the short-term demand pattern in the historical records to obtain the short-term demand features. A self-attention mechanism is introduced to weight the temporal features of each region, and the global features of the region are obtained by summing all the weighted local features of the region. In the long-term demand encoder, global features of the entire time series are extracted through convolution; In the event-related demand evolution encoder, LSTM is used to extract features, and the neural Hawkes process is introduced to simulate the event impact intensity that changes over time to perform demand pattern fusion.

3. The system for forecasting the demand for multiple categories of materials under public health emergencies according to claim 2, characterized in that: The feature extraction in the long-term demand encoder is specifically as follows: input the long-term demand sequence but in, Represents sequence characteristics, represents the required features, sigmoid(·) and tanh(·) are activation functions, Conv(·) is the temporal convolution operation, and the convolution kernel is k; w s 、w t , b are trainable parameters.

4. The system for forecasting the demand for multiple categories of materials under public health emergencies according to claim 2, characterized in that: The output of the timing requirement evolution encoding module is: in, Indicates different types of demand characteristics. represents the demand evolution characteristics, β represents the weight parameter of each scale encoder, Represents the weight parameter of the feature, s represents short-term features, l represents long-term features, and e represents event evolution features.

5. The system for forecasting the demand for multiple categories of materials under public health emergencies according to claim 1, characterized in that: The adaptive matrix is ​​calculated as follows: Among them, the softmax(·) function is used as an activation function to eliminate weak connections between categories; the ReLU(·) function is used as an activation function to normalize the adaptive adjacency matrix; E1 and E2 are vector representations of randomly initialized category pairs.

Images