Multimodal transport logistics scheme recommendation algorithm based on customer demands

Abstract

The invention provides a multimodal transport logistics scheme recommendation algorithm based on customer demands, and the algorithm employs a feature extraction module which is used for converting original data into a model processable form, extracting features related to a logistics transport scheme from the original data, and transmitting the features to a database; a composite neural network model module fusing an embedded layer, an attention mechanism, a cross network and a deep network is constructed, meanwhile, the module comprehensively considers factors such as comprehensive score, diversity and deduplication of model prediction so as to ensure high quality and practicability of recommendation results, a weighted Huber loss function is introduced, and the recommendation accuracy is improved. The prediction performance of transportation cost and time is optimized at the same time, sorting loss (ListMLE loss) is introduced to optimize the sorting performance of the model, the model accurately recognizes customer demands, a personalized scheme is generated, an efficient decision support tool is provided for coastal port logistics enterprises, the service quality is improved, the operation cost is reduced, and the competitiveness is enhanced. And intelligent logistics technology application and port logistics intelligent upgrading are promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of logistics solution recommendation algorithms, and in particular to a multimodal transport logistics solution recommendation algorithm based on customer demand. Background Art

[0002] With the acceleration of global economic integration, the connection between international trade and regional economies has become increasingly close. As a key node connecting global trade, port logistics has become increasingly important. Port logistics not only fulfills fundamental functions such as cargo transportation, warehousing, and distribution, but is also a core driver of regional economic development. Furthermore, my country's multimodal transport was first applied to foreign trade, primarily in economically developed regions such as the Yangtze River Delta, the Pearl River Delta, and the Bohai Rim, but its volume of transport remains relatively low.

[0003] However, with the rapid development of intermodal transport, the port logistics industry is facing new challenges. Customer demands for logistics services are becoming increasingly diverse and personalized. Customers are no longer satisfied with traditional, standardized logistics services, but are increasingly focused on efficient, flexible, and cost-effective logistics solutions. Traditional logistics solution decision-making systems, mostly based on experience or simple rule-based engines, struggle to adapt to complex and changing customer needs and are unable to provide accurate and personalized logistics solution recommendations. This not only limits the service quality and competitiveness of port logistics companies, but also restricts the in-depth application of smart logistics technologies in port intermodal transport scenarios.

[0004] Therefore, focusing on intelligent decision-making technology for customer-oriented logistics solutions and leveraging deep learning to construct an intelligent recommendation decision-making system that accurately captures customer needs and generates personalized logistics solutions is of great theoretical and practical significance. From a theoretical perspective, this research will enrich the theoretical framework of logistics demand analysis by deeply exploring and modeling customer demand characteristics, providing new perspectives and methodologies for subsequent related research. From a practical perspective, the findings of this research can provide port logistics companies with efficient and flexible decision-making support tools. Through this intelligent decision-making system, companies can quickly respond to customer needs and generate optimal logistics solutions, thereby improving service quality, reducing operating costs, and enhancing market competitiveness. Furthermore, this research will provide technical support for Lianyungang Port to develop into an international smart logistics hub, enabling it to play a greater role. Furthermore, the theoretical and technical findings of this research will serve as a reference for other port logistics companies nationwide, promoting the construction and development of smart logistics platforms and facilitating the intelligent upgrade of my country's port logistics industry.

[0005] Some existing research has attempted to address the logistics recommendation problem by improving deep learning models or algorithm design. For example, some researchers have proposed a two-stage vehicle-cargo matching recommendation system to address the distribution needs of large-scale goods and emergency bulk goods in modern logistics. However, research on logistics recommendation systems for intermodal transport scenarios is relatively scarce, and these methods may have certain limitations in intermodal transport scenarios. Summary of the Invention

[0006] In response to the shortcomings of existing technologies, the present invention proposes a multimodal transport logistics solution recommendation algorithm based on customer needs. By introducing deep learning technology to accurately capture customer demand characteristics, it can better handle the complexity brought about by the coordinated operation of multiple transportation modes in multimodal transport, and improve the intelligence level of the recommendation system. At the same time, the theoretical and technical achievements of this algorithm can provide reference and reference for other port logistics companies in the country, promote the construction of smart logistics platforms, and assist in the intelligent upgrade of my country's port logistics industry. The present invention proposes a multimodal transport logistics solution recommendation algorithm based on customer needs, which consists of the following steps:

[0007] S1: Use a feature extraction module to convert raw data into a form that can be processed by the model. Through a series of feature engineering methods, features related to logistics and transportation solutions are extracted from the raw data to meet customer needs and improve the model's predictive performance.

[0008] S2: Build a composite neural network model module that integrates embedding layers, attention mechanisms, cross networks, and deep networks to meet customers' dual needs for shipping cost and time.

[0009] S3: The recommendation generation module used can filter out the top K recommended options that best meet the customer's specific needs (such as origin, destination, and cargo type) from all possible transportation options. This module also takes into account factors such as the model's comprehensive score, diversity, and deduplication to ensure the high quality and practicality of the recommendation results.

[0010] S4: A weighted Huber loss function is introduced to simultaneously optimize the prediction performance of transportation cost and time, and a ranking loss (ListMLE loss) is introduced to optimize the ranking performance of the model.

[0011] Furthermore, the feature extraction module in S1 further includes the following features:

[0012] S1-1: First, feature selection is performed to filter out task-relevant and predictive features from the raw data. Based on the business requirements of the multimodal logistics recommendation system, the following features are selected: origin, destination, cargo type, transit station, total transport time, transport mode, transport time for each segment, transport mode for each segment, and standard amount.

[0013] S1-2: Next, feature construction is performed to generate new features from the original data, aiming to capture the hidden patterns and information in the data. The following new features are constructed: number of transfers and diversity of transportation modes;

[0014] S1-3: Feature encoding is the process of converting categorical features into numerical features so that the model can process them. The following encoding methods are mainly used: fixed encoding, transport mode encoding, and hash encoding;

[0015] S1-4: Feature standardization is the process of scaling the values of continuous features to a specific range to improve model training efficiency and performance. Standard Scaler is used to standardize continuous features.

[0016] Furthermore, the neural network module in S2 is a composite neural network model module that integrates an embedding layer, an attention mechanism, a cross network, and a deep network to meet customers' dual needs for transportation cost and time. Specifically, it includes the following:

[0017] S2-1: In order to effectively convert categorical features into numerical forms that can be processed by the model, an embedding layer is used. The embedding layer maps categorical features into a low-dimensional continuous vector space, which not only reduces the dimensionality of the features but also preserves the semantic information between the features.

[0018] S2-2: The attention mechanism uses weighted summation of input features to highlight important features while suppressing unimportant features. This mechanism enables the model to pay more attention to task-related features when processing sequence data and feature interactions.

[0019] S2-3: Use a cross-network module to learn feature interactions. The cross-network is particularly suitable for processing sparse features and high-dimensional features. It calculates feature interactions layer by layer, enhancing the model's ability to learn feature combinations.

[0020] S2-4: A deep network is a multi-layer perceptron (MLP) that performs nonlinear transformations on features and makes final predictions. It enhances the model's expressiveness through multiple fully connected layers and activation functions. Specifically, through multi-layer perceptrons and residual connections, deep networks can effectively handle complex nonlinear relationships, thereby improving the model's predictive performance.

[0021] S2-5: Fuse the output of the attention mechanism attn_out and the output of the cross network cross_out to form a comprehensive feature representation x fused , and then the fused features are used as the input of the deep network, and the two dimensions of the output represent the predicted values of transportation cost and time respectively.

[0022] Furthermore, the recommendation generation module used in S3 specifically includes the following:

[0023] S3-1: Filter eligible candidate transportation solutions from the data set based on the customer's origin, destination, and cargo type;

[0024] S3-2: Use the trained neural network model to predict the selected candidate options and generate the predicted transportation cost and time for each option. Based on these predicted values, calculate the comprehensive score for each option;

[0025] S3-3: To ensure the diversity of recommended solutions, a diversity scoring mechanism is introduced. The diversity score is evaluated by calculating the similarity between the new solution and the recommended solutions. The lower the similarity, the higher the diversity score.

[0026] S3-4: Combine the comprehensive score of each candidate solution with the diversity score to obtain the final recommendation score;

[0027] S3-5: Sort all candidate solutions by recommendation score and remove duplicate solutions. Finally, select the top K solutions as the recommended results and provide them to the customer.

[0028] Furthermore, the weighted Huber loss function is introduced in S4 to simultaneously optimize the prediction performance of transportation cost and time, and the ranking loss (ListMLE loss) is introduced to optimize the ranking performance of the model, specifically including the following:

[0029] S4-1: The Huber loss function used is a loss function that combines the mean square error (MSE) and the mean absolute error (MAE), which can be more robust to outliers;

[0030] S4-2: The specific calculation formula of the loss function is:

[0031]

[0032] S4-3: Where δ is a threshold used to control the conversion of the loss function between MSE and MAE. In this method, δ is set to 1.0;

[0033] S4-4: Since the model predicts both transportation cost and time, the total loss function is the weighted sum of the two tasks;

[0034] S4-5: The specific calculation formula of the total loss function is:

[0035] TotalLLoss=α·HuberLoss cost +(1-α)·HuberLoss time

[0036] S4-6: where α is a weight parameter used to balance the losses of the two tasks. In this method, α is set to 0.5, indicating that the prediction of transportation cost and time are given equal weight;

[0037] S4-7: To optimize the ranking performance of the model, the ListMLE loss function is introduced. The ListMLE loss function optimizes the ranking performance of the model by maximizing the likelihood of the predicted ranking;

[0038] S4-8: The specific calculation formula of the loss function is:

[0039]

[0040] S4-9: Where N is the number of samples, y pred,i is the model's predicted value for the i-th sample.

[0041] S4-10: The specific calculation formula of the total loss function is:

[0042] TotalLoss=TotalLoss Huber +β·ListMLELoss

[0043] Compared with the existing technology, this application proposal has the following main technical advantages:

[0044] This problem is addressed by developing an algorithm tailored to the complex requirements of recommending logistics solutions in multimodal transport scenarios. This approach incorporates deep learning technology to accurately capture customer demand characteristics, better handling the complexities inherent in the coordinated operation of multiple transport modes in multimodal transport, and enhancing the intelligence of the recommendation system. The theoretical and technical findings of this algorithm can also serve as a reference for other port logistics companies nationwide, promoting the development of smart logistics platforms and contributing to the intelligent upgrade of my country's port logistics industry.

[0045] Meeting domain needs: The proposed multimodal logistics solution recommendation algorithm based on customer needs can meet the complex needs of logistics solution recommendation in multimodal transport scenarios and is an algorithm suitable for this field.

[0046] Improving recommendation accuracy: This method introduces deep learning technology to accurately capture customer demand characteristics. It can better handle the complexity brought about by the coordinated operation of multiple transportation modes in multimodal transport, and improve the intelligence level of the recommendation system.

[0047] Reference value: The theoretical and technical achievements of this algorithm can provide reference and reference for other port logistics companies across the country, promote the construction of smart logistics platforms, and assist in the intelligent upgrade of my country's port logistics industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a diagram of the architecture of an algorithm model for recommending a multimodal transport logistics solution based on customer needs provided by the present invention;

[0049] Figure 2 This is a diagram showing the influence of different K values on the index values provided by the present invention;

[0050] Figure 3 This is a diagram showing the impact of different hidden layer dimensions on index values provided by the present invention;

[0051] Figure 4 This is a diagram showing the impact of different depth network layers on the index value provided by the present invention. DETAILED DESCRIPTION

[0052] The present invention is further described below with reference to the accompanying drawings and examples. However, the present invention can be implemented in many different ways and should not be construed as limited to the illustrated embodiments; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0053] Example 1

[0054] Dataset: The Lianyungang Port Intermodal Logistics Dataset. This dataset describes intermodal port logistics and can be used for intermodal logistics recommendations. The dataset consists of 3,600 training samples, 1,030 validation samples, and 514 test samples, for a total of 5,144 samples. This dataset contains information related to freight transportation and billing, and can be used to analyze transportation business processes and financial performance. The dataset contains 5,144 records, covering 67 different aspects. Order-related information in the dataset includes unique identifiers such as order number, waybill number, and bill number, which are used to distinguish between orders, waybills, and bills. Time information such as order placement time, order acceptance time, and transportation completion time is also included, which can be used to analyze the time span of business processes. Cargo information includes fields such as cargo type, empty / loaded container, quantity, container type, container dimensions, and cargo weight, describing basic cargo attributes. Location information includes origin and destination codes, as well as information on multiple transit stations. The dataset also contains the transportation methods and travel time between these locations, which can be used to analyze transportation routes and efficiency. Business participant information includes information such as the order acceptor, ordering company, ordering company, carrier, biller name, and service provider company name, clearly defining each participant in the business process. Invoice information includes invoice type, billing period, billing month, standard amount, floating discount point, floating amount, and total of other amounts, useful for financial accounting and analysis. Furthermore, the majority of transport records in the dataset involve three transit stations, with a few involving two or four.

[0055] In this embodiment, the evaluation indicators used include hit rate and normalized discounted cumulative gain, where the value of K is 5.

[0056] Hit-Ratio@K is used to measure whether the model contains the true optimal solution among the first K recommended solutions. Its calculation formula is as follows:

[0057]

[0058] In this formula, is the indicator function, Top-K represents the top K predicted values, N is the number of samples in the test set, and y i is the ordering of the real values, is the predicted value of the model.

[0059] Normalized Discounted Cumulative Gain (NDCG@K) is used to measure the quality and ranking performance of the model's recommendation solutions. Its calculation formula is:

[0060]

[0061] In this formula, y i is the ordering of the real values, is the predicted value of the model, and K is the length of the recommendation list.

[0062] The experimental environment for this example is as follows: processor: Intel Core i5-12600KF, memory: 32GB DDR5 (6400MHz), graphics card: NVIDIA GeForce RTX 4060Ti GPU, storage hard disk: 1024GB SSD, operating system: Windows 11. Development environment: Python version 3.10, PyTorch version 2.6.0, and torchvision version 0.21.0.

[0063] Table 1 Algorithm performance comparison

[0064]

[0065] By comparing model parameters, our COMLRA method achieves higher HR@5 and NDCG@5 than the comparison models LightGCN and HeCo, as shown in Table 1. LightGCN is a simplified graph convolutional network that removes unnecessary complex operations and focuses on the core neighborhood aggregation operation in GCN, thereby improving model training efficiency and recommendation performance. The HeCo model encodes nodes from both a network pattern view and a meta-path view, and introduces a graph mask mechanism, allowing these two views to complement and supervise each other, further improving model performance.

[0066] In summary, the present invention proposes a multimodal logistics solution recommendation algorithm based on customer demand. Previous research was mainly based on experience or simple rule engines, which made it difficult to adapt to complex and changeable customer needs and could not provide accurate and personalized logistics solution recommendations. Research on logistics recommendation systems for multimodal transport scenarios is still relatively scarce, and these methods may have certain limitations in multimodal transport scenarios. In order to solve this problem, a new multimodal logistics solution recommendation algorithm (COMLRA) based on customer demand is proposed. By introducing deep learning technology to accurately capture customer demand characteristics, it can better handle the complexity brought about by the collaborative operation of multiple transportation modes involved in multimodal transport, and improve the intelligence level of the recommendation system. At the same time, the theoretical and technical achievements of the algorithm can provide reference and reference for other port logistics companies across the country, promote the construction of smart logistics platforms, and assist in the intelligent upgrade of my country's port logistics industry.

[0067] The above embodiment merely represents one embodiment of the present invention. Although the description is relatively specific and detailed, it should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and such modifications and improvements shall fall within the scope of protection of the present invention.

Claims

1. A multimodal logistics solution recommendation algorithm based on customer needs, characterized by: The specific steps are as follows: S1: Use a feature extraction module to convert raw data into a form that can be processed by the model. Through a series of feature engineering methods, features related to logistics and transportation solutions are extracted from the raw data to meet customer needs and improve the model's predictive performance. S2: Build a composite neural network model module that integrates embedding layers, attention mechanisms, cross networks, and deep networks to meet customers' dual needs for shipping cost and time. S3: The recommendation generation module used can filter out the top K recommended options that best meet the customer's specific needs (such as origin, destination, and cargo type) from all possible transportation options. This module also takes into account factors such as the model's comprehensive score, diversity, and deduplication to ensure the high quality and practicality of the recommendation results. S4: A weighted Huber loss function is introduced to simultaneously optimize the prediction performance of transportation cost and time, and a ranking loss (ListMLE loss) is introduced to optimize the ranking performance of the model.

2. The multimodal transport logistics solution recommendation algorithm based on customer needs according to claim 1 is characterized in that: The feature extraction module in S1 further includes the following features: S1-1: First, feature selection is performed to filter out task-relevant and predictive features from the raw data. Based on the business requirements of the multimodal logistics recommendation system, the following features are selected: origin, destination, cargo type, transit station, total transport time, transport mode, transport time for each segment, transport mode for each segment, and standard amount. S1-2: Next, feature construction is performed to generate new features from the original data, aiming to capture the hidden patterns and information in the data. The following new features are constructed: number of transfers and diversity of transportation modes; S1-3: Feature encoding is the process of converting categorical features into numerical features so that the model can process them. The following encoding methods are mainly used: fixed encoding, transport mode encoding, and hash encoding; S1-4: Feature standardization is the process of scaling the values of continuous features to a specific range to improve the training efficiency and performance of the model. Standard Scaler is used to standardize continuous features.

3. The multimodal transport logistics solution recommendation algorithm based on customer demand according to claim 1 is characterized in that: The neural network module in S2 is a composite neural network model module that integrates an embedding layer, an attention mechanism, a cross network, and a deep network to meet customers' dual needs for shipping cost and time. Specifically, it includes the following: S2-1: To effectively convert categorical features into a numerical form that the model can process, an embedding layer is used. The embedding layer maps categorical features into a low-dimensional continuous vector space, which not only reduces the dimensionality of the features but also preserves the semantic information between the features. S2-2: The attention mechanism is used to perform weighted summation on the input features, highlighting important features while suppressing unimportant features. This mechanism enables the model to pay more attention to task-related features when processing sequence data and feature interactions. S2-3: Use the cross network module to learn feature interactions. The cross network is particularly suitable for processing sparse features and high-dimensional features. It enhances the model's ability to learn feature combinations by calculating feature interactions layer by layer. S2-4: A deep network is a multi-layer perceptron (MLP) that performs nonlinear transformations on features and makes final predictions. It uses multiple fully connected layers and activation functions to enhance the model's expressiveness. Specifically, through multi-layer perceptrons and residual connections, the deep network can effectively handle complex nonlinear relationships, thereby improving the model's predictive performance. S2-5: Fuse the output of the attention mechanism attn_out and the output of the cross network cross_out to form a comprehensive feature representation x fused , and then the fused features are used as the input of the deep network, and the two dimensions of the output represent the predicted values of transportation cost and time respectively.

4. The multimodal transport logistics solution recommendation algorithm based on customer demand according to claim 1, characterized in that: The recommendation generation module used in S3 specifically includes the following: S3-1: Filter eligible candidate transportation solutions from the data set based on the customer's origin, destination, and cargo type; S3-2: Use the trained neural network model to predict the selected candidate options, generate predicted values for transportation cost and time for each option, and calculate the comprehensive score for each option based on these predicted values; S3-3: To ensure the diversity of recommended solutions, a diversity scoring mechanism is introduced. The diversity score is evaluated by calculating the similarity between the new solution and the recommended solutions. The lower the similarity, the higher the diversity score. S3-4: Combine the comprehensive score of each candidate solution with the diversity score to obtain the final recommendation score; S3-5: Sort all candidate solutions by recommendation score and remove duplicate solutions. Finally, select the top K solutions as the recommendation results and provide them to the customer.

5. The multimodal transport logistics solution recommendation algorithm based on customer demand according to claim 1, characterized in that: S4 introduces a weighted Huber loss function to simultaneously optimize the prediction performance of transportation cost and time, and introduces a ranking loss (ListMLE loss) to optimize the ranking performance of the model. Specifically, it includes the following: S4-1: The Huber loss function used is a loss function that combines the mean square error (MSE) and the mean absolute error (MAE), which can be more robust to outliers; S4-2: The specific calculation formula of the loss function is: S4-3: where δ is a threshold used to control the conversion of the loss function between MSE and MAE. In this method, δ is set to 1.

0. S4-4: Since the model predicts both transportation cost and time, the total loss function is the weighted sum of the two tasks; S4-5: The specific calculation formula of the total loss function is: TotaLLoss=α·HuberLoss cost +(1-α)·HuberLoss time S4-6: where α is a weight parameter used to balance the losses of the two tasks. In this method, α is set to 0.5, indicating that the prediction of transportation cost and time is given equal importance; S4-7: In order to optimize the ranking performance of the model, the ListMLE loss function is introduced. The ListMLE loss function optimizes the ranking performance of the model by maximizing the likelihood of the predicted ranking; S4-8: The specific calculation formula of the loss function is: S4-9: Where N is the number of samples, y pred,i is the model's predicted value for the i-th sample; S4-10: The specific calculation formula of the total loss function is: TotalLoss=TotalLoss Huber +β·ListMLELoss。

Images