KR102591504B1 - Logistics prediction technology using product association graph structure - Google Patents

Abstract

The logistics prediction method performed by the logistics prediction system includes deriving correlation rules between products through correlation analysis based on transaction history information of the products; generating a graph structure based on the association rules between the derived products; Outputting an embedding vector of a target product reflecting the correlation according to change information between each product and neighboring products using the generated graph structure; And it may include predicting the future logistics volume for the target product from the output embedding vector of the target product using a multi-neural network.

Description

Logistics prediction technology using product correlation graph structure {LOGISTICS PREDICTION TECHNOLOGY USING PRODUCT ASSOCIATION GRAPH STRUCTURE}

The explanation below is about logistics forecasting technology.

As the e-commerce market grew rapidly, the logistics industry also grew together, but the complexity of corporate supply chains and consumers' demands for logistics quality became more sophisticated, making it difficult for individual companies to acquire expertise in the logistics field. In order to satisfy these consumer demands, the logistics industry also needs a more efficient and accurate logistics forecasting method than before.

Technologies for logistics prediction include predicting the logistics of products based on factors including time series information, or predicting for each product based on sequential logistics information by assuming the relationship between products independently. A technology for predicting logistics based on factors including time series information is to use logistics information prior to product prediction and factors adding time characteristics as input to a machine learning model such as SVM (Support vector machine). There is a prediction technology, and a technology that independently assumes the relationship between products extracts the sequential information of each product from transaction history information, and uses RNN (Recurrent neural network) or LSTM for the sequential information of each extracted product. There is a technology to predict future logistics based on long short-term memory.

These prior technologies predict logistics by adding time series factors or make independent logistics predictions based on sequential information. In the case of a method of predicting logistics by adding time series factors, the logistics information for the product in question uses only recent information, so past logistics information depends on factors with added time characteristics, which greatly affects how factors are created. Not only does this result in loss of information, but it also causes information loss, making logistics prediction difficult. In the case of the method of predicting independent logistics based on sequential information, it is performed based on sequential information, but assumes an independent relationship. However, it is greatly influenced by abnormal values in the logistics information of the product, resulting in large errors, and also makes the products independent. However, it is difficult to make accurate predictions because deep relationship information is not sufficiently reflected between products that appear in the actual purchase stage.

A method and system for predicting future logistics volume by creating a graph structure based on correlation analysis to reflect the relationship information of each product based on the transaction history information of the products and sequentially extracting the transaction information of the products using deep learning technology. can be provided.

The logistics prediction method performed by the logistics prediction system includes deriving correlation rules between products through correlation analysis based on transaction history information of the products; generating a graph structure based on the association rules between the derived products; Outputting an embedding vector of a target product reflecting the correlation according to change information between each product and neighboring products using the generated graph structure; And it may include predicting the future logistics volume for the target product from the output embedding vector of the target product using a multi-neural network.

The step of deriving the correlation rule is to derive a correlation rule between products including the reliability and improvement between each product using transaction history information for a specific period of the products through product information included in the receipt information. It may include steps.

In the step of generating the graph, in order to reflect the relationship of change information over time between the target product and neighboring products in the graph, the identification information of each product is corresponded to a node, and the reliability and improvement included in the correlation rule is performed. It may include the step of expressing the connection strength of the edge connecting the nodes using degrees.

The reliability refers to the degree to which users who purchased the target product also purchased neighboring products, and the improvement refers to the degree to which users who purchased the target product also purchased neighboring products compared to neighboring products. The connection strength of the edge connecting can be derived through the product of the reliability and the improvement.

The step of outputting the embedding vector of the target product may include extracting feature information using a graph network and a neural network model by applying temporal logistics change information over a preset period in the generated graph structure. .

The step of outputting the embedding vector of the target product includes outputting an embedding vector that reflects the correlation according to the sales volume change information of each product and neighboring products as the generated graph structure is processed with a deep learning-based graph network. It can be included.

In the step of outputting the embedding vector of the target product, the change in sales volume of each product is analyzed as the daily sales volume of each product for a preset period for each node configured in the graph network is input to the neural network structure, and the change in sales volume of each product is analyzed. The node information that analyzed the change in sales volume may be processed based on the edges of the graph structure and integrated into the embedding vector of the target product.

The multi-neural network may be constructed to consist of a plurality of neural networks according to the length to be predicted, extract feature information from the output embedding vector of the target product, and predict sales volume based on the extracted feature information.

The step of predicting the future logistics volume for the target product may include predicting the future logistics volume for the target product during a period corresponding to the number of a plurality of neural networks configured in the multi-neural network.

The step of predicting the future logistics volume for the target product involves predicting the future logistics volume for the target product using feature information extracted from the embedding vector in a first neural network (t+1) among a plurality of neural networks configured in the multi-neural network. It may include a prediction step.

The step of predicting the future logistics volume for the target product includes feature information extracted from the embedding vector in the second neural network (t+2) among the plurality of neural networks configured in the multi-neural network and the first neural network (T+1). It may include predicting the future logistics volume for the target product using characteristic information.

The step of predicting the future logistics volume for the target product may include providing a comparison result of comparing the predicted future logistics volume and the actual logistics volume in a graph.

It may include a computer program stored in a non-transitory computer-readable recording medium to execute the logistics prediction method on the logistics prediction system.

The logistics prediction system includes a correlation rule derivation unit that derives correlation rules between products through correlation analysis based on transaction history information of the products; a graph structure generator that generates a graph structure based on the association rules between the derived products; a past logistics information embedding unit that outputs an embedding vector of the target product reflecting the correlation according to change information between each product and neighboring products using the generated graph structure; and a future logistics prediction unit that predicts the future logistics volume for the target product from the output embedding vector of the target product using a multi-neural network.

Because product logistics are predicted based on a graph structure created based on correlation analysis, accuracy can be improved over prediction technologies based on existing independent relationships.

By deriving correlation rules based on transaction history information, it is possible to identify the relationships and trends of the latest products. By structuring a graph based on correlation analysis and reflecting it in future logistics predictions, predictions can be made taking into account the relationships between the target product and neighboring products. make it possible In addition, since predictions are made based on change information of neighboring products, they are robust to abnormal values such as noise that appear in the target product information, enabling more accurate and stable predictions.

1 is a diagram for explaining the operation of a logistics prediction system according to an embodiment.
Figure 2 is a diagram for explaining an operation of deriving a correlation rule in a logistics prediction system according to an embodiment.
Figure 3 is a diagram for explaining an operation of generating a graph in a logistics prediction system according to an embodiment.
Figure 4 is a diagram for explaining an operation of embedding past logistics information in a logistics prediction system according to an embodiment.
Figure 5 is a diagram for explaining an operation of predicting future logistics information in a logistics prediction system according to an embodiment.
Figure 6 is a block diagram for explaining the configuration of a logistics prediction system according to an embodiment.
Figure 7 is a flowchart for explaining a logistics prediction method in a logistics prediction system according to an embodiment.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

1 is a diagram for explaining the operation of a logistics prediction system according to an embodiment.

The logistics prediction system can perform the correlation rule derivation (110) process, graph structure creation (120) process, past logistics information embedding (130) process, and future logistics prediction (140) process. The logistics prediction system can generate a graph by estimating the relationship between products from transaction history information and process a logistics prediction algorithm that considers the relationship between products from the generated graph.

In the process of deriving correlation rules (110), the logistics prediction system can derive correlation rules between products through correlation analysis based on transaction history information of the products. The transaction history information of products may refer to information on the transaction history of the past (for example, the last three months) based on the present. Correlation analysis refers to deriving meaningful relationships between products by finding relationships within the data, that is, mutual or dependent relationships between products. The logistics prediction system can receive input data including customer purchase history information. At this time, the input data may include information on products purchased together according to receipt information classified by ID in the customers' purchase history information. The logistics prediction system can use products purchased using receipt information as input data to create association rules between products based on transaction history information for the past s months.

Referring to Figure 2, it is a diagram for explaining the operation of deriving an association rule. The logistics prediction system can derive correlation rules by performing correlation analysis based on transaction history information for the most recent s (s is a natural number) months. For example, there may be transaction history information for 3 months from the present. At this time, the transaction history information may include customer identification information (id), receipt identification information (id), and product identification information (id). Based on this transaction history information, an association rule including rule length, target product ID, neighboring product ID, reliability, and improvement can be derived.

In the graph structure creation process 120, the logistics prediction system may generate a graph structure based on correlation rules. The logistics prediction system can create a graph structure based on correlation rules derived based on past transaction history information. Referring to FIG. 3, it is a diagram for explaining the operation of generating a graph. The graph structure is a structure used in deep learning models to analyze based on the connectivity between each object, allowing the correlation between each product and neighboring products to be analyzed more efficiently. Nodes in the graph structure correspond to the ID information of each product, and the edges connected between nodes can be connected or unconnected depending on the association rule of each node, and the reliability and improvement of the association rule are used to connect nodes. It can be defined as the connection strength of the edge. For example, the connection strength of edges connecting nodes can be derived through the product of the reliability of the association rule and the degree of improvement. The logistics prediction system can reflect the relationship between change information (e.g., change in logistics volume, change in sales volume, etc.) between the target product and neighboring products over time through a graph structure.

In detail, Table 1 shows the correlation analysis indicators.

Table 1:

Logistics prediction systems can derive the significance of rules through correlation analysis indicators. Association rules may include confidence and lift. Reliability refers to the degree to which users who purchased the target product also purchased neighboring products (P(Neighbor|Target). Improvement refers to an indicator that judges the reliability of the rule (P(Neighbor|Target)/P(Neighbor )). At this time, if the improvement value is 1 or more, it can be judged to be reliable.

In the process of past logistics information embedding (130), the logistics prediction system applies temporal changes in logistics information for the last t (t is a natural number) days to the graph structure created in the graph structure generation (120) process, forming a graph network and neural network. Feature information can be extracted based on the model. At this time, meaningful feature information for logistics prediction can be extracted without target data through graph networks and neural network models.

The logistics prediction system can extract information on the change in logistics volume of a product over time by using the neural network structure of past logistics information for t days, and extract information on the change in logistics volume of the predicted target product and change information on neighboring products connected in a graph structure. It can be provided as input data to the graph network. Referring to FIG. 4, it is a diagram for explaining the operation of embedding past logistics information. Through the past logistics information embedding process, meaningful change information can be estimated by updating the information of the target product through change information of neighboring products. The graph structure created in the graph structure creation process (120) can be processed with a deep learning-based graph network to output an embedding vector that reflects the correlation according to the change in sales volume of each product and neighboring products. At each node of the graph network, the daily sales volume of each product for the most recent t days can be input as input data to the neural network structure. Through the neural network structure, changes in daily sales volume of each product over the last t days can be analyzed. Node information that analyzes the daily sales volume of each product can be processed based on the edges connected in the graph network and integrated into the embedding vector of the target product. The embedding vector of the target product output from the graph network is information that comprehensively analyzes changes in sales volume of the target product and neighboring products and the correlation between them, focusing on the target product, and can be used to predict future logistics.

In the process of predicting future logistics 140, the logistics prediction system can predict future logistics volume by reflecting change information based on relationship information between the target product and neighboring products according to a prediction algorithm. The logistics prediction system processes the embedding vector output through the graph network in the process of past logistics information embedding (130) with a multi-neural network structure to reflect the future sales volume of the target product for H (H is a natural number) days, reflecting the correlation with neighboring products. can be predicted. The logistics prediction system can predict the future logistics volume of the target product by reflecting change information based on the correlation between the target product and neighboring products according to a prediction algorithm. Referring to Figure 5, this is a diagram to explain the operation of predicting future logistics information. In the embodiment, the operation of predicting future logistics for day H through a multi-neural network structure will be explained as an example.

Multi-neural network is a neural network structure for more accurately predicting multiple viewpoints that are sequentially predicted. The structure of the multi-neural network consists of H (H is a natural number) neural networks depending on the length to be predicted, and each neural network structure configured in the multi-neural network can be configured to perform two steps. In the first step, feature information is extracted from the embedding vector, and in the second step, sales volume can be predicted based on the extracted feature information. Among the H neural networks, the t+1th neural network predicts future logistics using feature information extracted from only the embedding vector, and from the t+2th neural network onwards, feature information extracted from the t+2 neural network and embeddings from the t+1 neural network are used. Feature information extracted from the vector may also be provided as input data. At this time, the feature information extracted from the previous neural network based on the additionally provided current neural network helps extract meaningful feature information from the embedding vector of the previous neural network so that future logistics can be predicted more accurately from the feature information in the second step. do. In this way, the feature information used to predict the previous state information is used together when predicting the next state information, thereby performing dependent prediction rather than independent prediction. In addition, in the process of future logistics prediction 140 through the above-mentioned operation, the logistics prediction system can output the logistics forecast value for H days of the target product. The logistics forecasting system can perform time series forecasting by reflecting the correlation of target products. As the logistics prediction system predicts future logistics volume, it can display the actual logistics volume and predicted logistics volume in a graph.

The results according to the comparison model and quantitative indicators to verify the significance of the technology proposed in the example are shown. Table 2 shows the quantitative results of the model.

Table 2:

The indicator used is RMSE, which is an indicator that shows the difference between predicted demand and actual demand.

Benchmark models were those that were widely used in the time series field (LSTM, 1D-CNN) and models that only used multi MLP among the proposed techniques were selected, and the selected models were compared with the proposed method. LSTM, which has shown high performance in past time series data, and 1D-CNN, which has recently been actively studied in the stock field, can be compared.

According to the embodiment, the relationship between each product is approached as an independent relationship and independent logistics prediction is performed without considering the relationship, or logistics is predicted by creating meaningful variables for each product based on past transaction history information. Therefore, it is possible to solve the problem of conventional logistics prediction technology that the accuracy of logistics prediction is low because the relationship between each product is assumed independently and the correlation between logistics of multiple products cannot be considered.

According to the embodiment, correlation analysis is performed based on transaction history information of products to structure the correlation between products into a graph, and logistics prediction is made after d (d is a natural number) days using a graph artificial neural network structure such as a graph network. By performing , it is possible to improve logistics prediction performance by considering the dependent relationship between the logistics volumes of products, utilizing transaction information of neighboring products, and reflecting the correlation with the target product.

FIG. 6 is a block diagram for explaining the configuration of a logistics prediction system according to an embodiment, and FIG. 7 is a flowchart for explaining a logistics prediction method in the logistics prediction system according to an embodiment.

The processor of the logistics prediction system 100 may include a correlation rule derivation unit 610, a graph structure generation unit 620, a past logistics information embedding unit 630, and a future logistics prediction unit 640. These processor components may be expressions of different functions performed by the processor according to control instructions provided by program codes stored in the logistics prediction system. The processor and its components may control the logistics prediction system to perform steps 710 to 740 included in the logistics prediction method of FIG. 7 . At this time, the processor and its components may be implemented to execute instructions according to the code of an operating system included in the memory and the code of at least one program.

The processor may load program code stored in a file of a program for a logistics prediction method into memory. For example, when a program is executed in a logistics prediction system, the processor can control the logistics prediction system to load program code from a program file into memory under the control of the operating system. At this time, the correlation rule derivation unit 610, graph structure generation unit 620, past logistics information embedding unit 630, and future logistics prediction unit 640 each execute instructions of the corresponding portion of the program code loaded in the memory. Accordingly, they may be different functional expressions of the processor for executing the subsequent steps 710 to 740.

In step 710, the correlation rule derivation unit 610 may derive correlation rules between products through correlation analysis based on transaction history information of the products. The correlation rule derivation unit 610 can use the transaction history information of the products for a specific period through the product information included in the receipt information to derive association rules between products, including the reliability and degree of improvement between each product. there is.

In step 720, the graph structure generator 620 may generate a graph structure based on the association rules between derived products. The graph structure generator 620 corresponds the identification information of each product to a node in order to reflect the relationship of change information over time between the target product and neighboring products in the graph, and the reliability and improvement levels included in the correlation rule. It can be expressed as the connection strength of the edges connecting nodes. At this time, reliability refers to the degree to which users who purchased the target product also purchased neighboring products, and improvement refers to the degree to which users who purchased the target product also purchased neighboring products compared to neighboring products. The connection strength of the connecting edge can be derived through the product of reliability and improvement.

In step 730, the past logistics information embedding unit 630 may output an embedding vector of the target product that reflects the correlation according to change information of each product and neighboring products using the generated graph structure. The past logistics information embedding unit 630 may apply temporal logistics change information over a preset period to the generated graph structure and extract feature information using a graph network and neural network model. As the generated graph structure is processed with a deep learning-based graph network, the past logistics information embedding unit 630 can output an embedding vector that reflects the correlation according to the sales volume change information of each product and neighboring products. The past logistics information embedding unit 630 analyzes the change in sales volume of each product as the daily sales volume of each product for a preset period for each node configured in the graph network is input into the neural network structure, and analyzes the change in sales volume of each product. One node information can be processed based on the edges of the graph structure and integrated into the embedding vector of the target product.

In step 740, the future logistics prediction unit 640 may predict the future logistics volume for the target product from the output embedding vector of the target product using a multi-neural network. At this time, the multi-neural network may be constructed to consist of a plurality of neural networks according to the length to be predicted, extract feature information from the output embedding vector of the target product, and predict sales volume based on the extracted feature information. The future logistics prediction unit 640 can predict the future logistics volume for the target product during a period corresponding to the number of plural neural networks configured in the multi-neural network. The future logistics prediction unit 640 can predict the future logistics volume for the target product using feature information extracted from the embedding vector in the first neural network (t+1) among the plurality of neural networks configured in the multi-neural network. The future logistics prediction unit 640 uses the feature information extracted from the embedding vector in the second neural network (t+2) and the feature information in the first neural network (T+1) among the plurality of neural networks configured in the multi-neural network to target the product. Future logistics volume can be predicted. The future logistics prediction unit 640 can provide a graphical comparison result comparing the future logistics volume predicted through a multi-neural network and the actual logistics volume.

The device described above may be implemented with hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), etc. , may be implemented using one or more general-purpose or special-purpose computers, such as a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. A processing device may execute an operating system (OS) and one or more software applications that run on the operating system. Additionally, a processing device may access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, a single processing device may be described as being used; however, those skilled in the art will understand that a processing device includes multiple processing elements and/or multiple types of processing elements. It can be seen that it may include. For example, a processing device may include a plurality of processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are possible.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. Software and/or data may be used on any type of machine, component, physical device, virtual equipment, computer storage medium or device to be interpreted by or to provide instructions or data to a processing device. It can be embodied in . Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the medium may be specially designed and configured for the embodiment or may be known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -Includes optical media (magneto-optical media) and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

As described above, although the embodiments have been described with limited examples and drawings, various modifications and variations can be made by those skilled in the art from the above description. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

Therefore, other implementations, other embodiments, and equivalents of the claims also fall within the scope of the claims described below.

Claims (14)

In the logistics prediction method performed by the logistics prediction system,
Deriving correlation rules between products through correlation analysis based on transaction history information of the products;
generating a graph structure based on the association rules between the derived products;
Outputting an embedding vector of a target product reflecting the correlation according to change information between each product and neighboring products through a graph network and neural network model using the generated graph structure; and
Predicting future logistics volume for the target product from the output embedding vector of the target product using a multi-neural network
Including,
The step of predicting the future logistics volume for the target product is,
Among the plurality of neural networks configured in the multi-neural network, a first neural network (t+1) predicts the future logistics volume for the target product using feature information extracted from the embedding vector, and predicts the future logistics volume for the target product, and Predicting the future logistics volume for the target product using feature information extracted from the embedding vector in the second neural network (t+2) among the neural networks and feature information in the first neural network (t+1)
Logistics forecasting method including. According to paragraph 1,
The step of deriving the association rule is,
A step of deriving association rules between products, including reliability and improvement between each product, using transaction history information for a specific period of time through product information included in receipt information.
Logistics forecasting method including. According to paragraph 1,
The step of generating the graph is,
In order to reflect the relationship of change information over time between the target product and neighboring products in the graph, the identification information of each product is mapped to a node, and the nodes are connected using the reliability and improvement included in the correlation rule. A step expressed in terms of the connection strength of the edge
Logistics forecasting method including. According to paragraph 3,
The reliability refers to the degree to which users who purchased the target product also purchased neighboring products,
The degree of improvement refers to the degree to which users who purchased the target product compared to neighboring products also purchased neighboring products,
A logistics prediction method, characterized in that the connection strength of the edge connecting the nodes is derived through the product of the reliability and the improvement. According to paragraph 1,
The step of outputting the embedding vector of the target product is,
Extracting feature information using a graph network and neural network model by applying temporal logistics change information over a preset period from the generated graph structure.
Logistics forecasting method including. According to clause 5,
The step of outputting the embedding vector of the target product is,
As the generated graph structure is processed with a deep learning-based graph network, outputting an embedding vector reflecting the correlation according to the sales volume change information of each product and neighboring products.
Logistics forecasting method including. According to clause 6,
The step of outputting the embedding vector of the target product is,
For each node configured in the graph network, the change in sales volume of each product is analyzed as the daily sales volume of each product during a preset period is input into the neural network structure, and the node information analyzing the change in sales volume of each product is included in the graph structure. A step of processing based on edges and integrating them into the embedding vector of the target product
Logistics forecasting method including. According to paragraph 1,
The multi-neural network consists of a plurality of neural networks according to the length to be predicted, extracts feature information from the output embedding vector of the target product, and is constructed to predict sales volume based on the extracted feature information. Logistics forecasting method. According to paragraph 1,
The step of predicting the future logistics volume for the target product is,
Predicting future logistics volume for the target product during a period corresponding to the number of plural neural networks configured in the multi-neural network
Logistics forecasting method including. delete delete According to clause 9,
The step of predicting the future logistics volume for the target product is,
Step of providing a graph comparing the predicted future logistics volume and actual logistics volume
Logistics forecasting method including. A computer program stored in a non-transitory computer-readable recording medium for executing the logistics prediction method of any one of claims 1 to 9 and 12 in the logistics prediction system. In the logistics prediction system,
a correlation rule derivation unit that derives correlation rules between products through correlation analysis based on transaction history information of the products;
a graph structure generator that generates a graph structure based on the association rules between the derived products;
a past logistics information embedding unit that outputs an embedding vector of the target product reflecting the correlation according to change information between each product and neighboring products through a graph network and neural network model using the generated graph structure; and
Future logistics prediction unit that predicts the future logistics volume for the target product from the output embedding vector of the target product using a multi-neural network
Including,
The future logistics forecasting department,
Among the plurality of neural networks configured in the multi-neural network, a first neural network (t+1) predicts the future logistics volume for the target product using feature information extracted from the embedding vector, and predicts the future logistics volume for the target product, and Among the neural networks, the second neural network (t+2) uses the feature information extracted from the embedding vector and the feature information of the first neural network (t+1) to predict the future logistics volume for the target product.
Logistics forecasting system.

Images