River-sea combined transportation path recommendation method based on user preference identification

Abstract

The invention relates to a river-sea combined transportation path recommendation method based on user preference identification, and the method comprises the steps: S1, collecting questionnaire results of a user for cost, time and carbon emission, and respectively generating a score vector and an AHP calculation weight vector; s2, fusing the score vector and the AHP calculation weight vector to obtain a first three-target preference weight; s3, dynamically updating a second three-target preference weight in combination with the first three-target preference weight based on historical orders and behavior data of the user; s4, fusing the first three-target preference weight and the second three-target preference weight to obtain a third three-target preference weight; and S5, on the basis of the third three-target preference weight, a total target function value of each candidate combined transport path is calculated in combination with a pre-configured total target function, and the candidate combined transport paths are sorted from optimal to inferior according to the total target function values. Compared with the prior art, the method has the advantages of remarkably improving the exclusive matching degree of the recommendation scheme and the like.

Description

Technical Field

[0001] This invention relates to the field of transportation, and in particular to a method for recommending river-sea intermodal transport routes based on user preference recognition. Background Technology

[0002] With the continuous advancement of the national "dual-carbon" strategy and multimodal transport development policies, river-sea intermodal transport, as an important mode of transportation connecting inland and sea routes, is playing an increasingly significant role in reducing logistics costs, improving transportation efficiency, and reducing carbon emissions. However, due to the involvement of multiple modes of transport and complex port node selection and route combinations, users face problems such as information asymmetry and inaccurate matching in the actual transportation plan selection, which restricts the personalization and intelligence of river-sea intermodal transport services.

[0003] Currently, most river-sea intermodal transport route recommendation systems mainly rely on static indicators such as cargo type, transportation cost, or transportation time for rule matching. They lack dynamic identification and fine-grained response to multi-dimensional personalized needs such as user subjective preferences, industry background, and environmental protection demands. This results in recommendation results that are highly generalized but weakly specific, making it difficult to meet the precise matching needs of high-value users, green transportation companies, or time-sensitive goods.

[0004] In practical applications, different types of users often exhibit significantly different weighting preferences among transportation costs, time efficiency, and carbon emissions. For example, manufacturing companies may prioritize cost control, while cold chain logistics companies place greater emphasis on delivery time, and environmental organizations tend to favor green and low-carbon routes. Existing recommendation models generally employ uniform scoring standards or static weight settings, which fail to reflect users' true preferences and can easily lead to significant biases in recommended solutions and low user satisfaction.

[0005] Therefore, there is an urgent need to build a river-sea intermodal transport route recommendation system that can identify individual user preferences and dynamically generate multi-objective weights accordingly. Summary of the Invention

[0006] The purpose of this invention is to provide a method for recommending river-sea intermodal transport routes based on user preference recognition.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for recommending river-sea intermodal transport routes based on user preference identification includes: Step S1: Collect user questionnaire results on cost, time and carbon emissions, and generate a scoring vector and an AHP-calculated weight vector respectively; Step S2: Fuse the rating vector and the AHP-calculated weight vector to obtain the first and third objective preference weights; Step S3: Based on the user's historical order and behavior data, dynamically update the second and third target preference weights in combination with the first and third target preference weights; Step S4: Combine the first and third objective preference weights and the second and third objective preference weights to obtain the third and third objective preference weights; Step S5: Based on the third objective preference weights, calculate the overall objective function value of each candidate intermodal transport route in combination with the pre-configured overall objective function, and sort the candidate intermodal transport routes from best to worst according to the overall objective function value.

[0008] The scoring vector is: T01 = ( , , ) in: For cost rating, A score for time. For carbon emissions scoring, T01 is the scoring vector. Questionnaires that directly score cost variables Questionnaires with time variables were scored directly. The questionnaire directly scores the carbon emission variable. The relative score of the cost variable compared to the time variable. The relative score of the time variable compared to the carbon emission variable. The relative scores of cost variables compared to carbon emission variables. , , , , , For pre-configured scoring weights.

[0009] The process of constructing the weight vector for AHP calculation includes: Step S1-1: Construct the AHP judgment matrix A: Where: 'a' is the importance evaluation value of cost relative to time, and its value is [value missing]. 'b' represents the importance assessment value of cost relative to carbon emissions, and its value is [value missing]. c represents the importance of time relative to carbon emissions, and its value is [value missing]. ; Step S1-2: Calculate the original weights of each objective based on the AHP judgment matrix A as the first and third objective preference weights.

[0010] in: The original weights for the i-th objective are: cost, time, and carbon emissions. Let be the geometric mean of the i-th target. The element in the i-th row and j-th column of the AHP judgment matrix; Step S2 includes Step S2-1: Calculate the corrected weights for each objective based on their original weights. in: The corrected weight for the i-th objective. This refers to the subjective rating weighting coefficient. Step S2-2: Normalize the corrected weights of each objective to obtain the normalized corrected weights of each objective as the preference weights of the second and third objectives.

[0011] Step S3 is implemented through a first weight prediction model, which is a Multi-Task Learning policy model based on the XGBoost core. In the structure of the first weighted prediction model, the first K trees form a shared layer, learning common features between tasks. Subsequently, each target is assigned K independent regression trees to learn their respective differences, ultimately outputting three independent weight values. , , ; The first weighted prediction model uses a multi-task joint loss function: in: For multi-task joint loss function, Let N be the label value of the cost item weight for the i-th task, and N be the number of tasks. The model prediction value for the cost item weights of the i-th task. Let be the label value of the time item weight for the i-th task. The model prediction value for the time term weight of the i-th task. Let the label value be the weight of the carbon emission item in the i-th task. Let be the model-predicted weights of the carbon emission items for the i-th task. This is a complexity regularization term used to prevent overfitting.

[0012] In step S4, the third three-target preference weight is the weighted sum of the first three-target preference weight and the second three-target preference weight, wherein the weight of the second three-target preference weight is the fusion coefficient γ, and the weight of the first three-target preference weight is 1-γ. The fusion coefficient γ is: in: To objectively learn the trust coefficient adjustment ratio, To calculate the consistency between user profiles and current model predictions by using structured modeling based on user basic attributes, industry characteristics, historical orders, and questionnaire results, Adjust the data volume / coverage ratio. Scoring the quality of behavioral data The confidence coefficient is based on the prediction.

[0013] Step S5 includes: Step S5-1: Construct a directed network containing port nodes, path edges, and transportation modes. Set the following attribute for each edge: path distance For transportation speed Unit transportation cost Unit carbon emission factor throughput per unit time Path set ; Step S5-2: Construct the order structure model and define each order as... ,in: As the starting and ending points, For the type of goods, For transport weight; Let the indicator preference weight vector satisfy... ; Step S5-3: Make basic assumptions and establish constraints for the order structure model to ensure the model's rationality and solvability. The basic assumptions include: (1) Orders can arrive at different times, and the information is known in advance, which belongs to static or quasi-dynamic scheduling. (2) The loading and unloading capacity and transportation time of ports and routes can be quantitatively estimated and vary with the hour. (3) The costs, emissions, and speeds of each mode of transportation are known or predictable fixed values. (4) The attributes and port service levels of each route segment do not fluctuate due to changes in orders; The constraints include transportation time window constraints, route capacity constraints, scheduling priority constraints, loading and unloading operation time constraints, and route continuity and feasibility constraints. The transportation time window constraint is as follows: in: For order O k The actual arrival time For order O k The earliest arrival time, For order O k Latest arrival time; The path capacity constraint is: in: This represents the throughput per unit time of the path edge or port node. This represents the maximum throughput per unit time on the path segment from node i to node j. Step S5-4: Generate multiple candidate intermodal routes for all orders based on order type, wherein the order type includes single order, non-competitive multiple orders, and competitive multiple orders; Step S5-5: Calculate the overall objective function value of each candidate intermodal transport route based on the pre-configured initial overall objective function, and sort the candidate intermodal transport routes from best to worst according to the overall objective function value.

[0014] Consider order sets Each order The quantity of goods is The candidate path set is P; for each path The process of calculating the overall objective function value and constructing the pre-configured initial overall objective function includes: Calculate the cost function: in: For cost function, For order k in path segment e i On the transportation cost, For transportation time; Calculate the time function in: It is a time function. This represents the transportation distance of the route segment. The average speed corresponding to the mode of transportation. These are the loading and unloading times, This refers to the waiting time caused by scheduling and queuing. Calculate the carbon emission function: in: For emission functions, Carbon emissions from the route segment; Construct a pre-configured initial overall objective function: in: , , These are the weights of the cost function, time function, and carbon emission function, respectively, and the sum of the three is 1.

[0015] In step S5-5, modified objective functions are generated based on the pre-configured initial overall objective function, for single orders, multiple orders without competition, and multiple orders with competition, which are modeled independently. The modified objective function for a single order considers only one order O. k The transportation process along path p∈P does not involve throughput competition. Calculations are performed using unit time, cost, and carbon emissions. The model is obtained as follows: in: For unit transportation cost, This represents the normalized unit transportation time. Normalized carbon emissions per unit of transportation This represents the minimum unit transportation cost among all candidate route options. This represents the maximum unit transportation cost among all candidate route options. Normalized unit transportation cost The modified objective function for multiple orders without competition is: in: The total shipping cost for a standardized, non-competitive multi-order shipment. For the standardized total time of multiple orders without competition, The total carbon emissions of multiple orders without competition after standardization.

[0016] A river-sea intermodal transport route recommendation device based on user preference recognition includes a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the method described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Accurately identify users' personalized preferences. Quantify users' subjective preferences in three categories of objectives—transportation costs, time efficiency, and carbon emissions—through questionnaires and AHP methods. Combine this with historical order behavior to dynamically depict the real needs of different users in different task scenarios.

[0018] 2. To address the cold start and data sparsity issues, when historical data is lacking, the system is primarily driven by subjective preferences, supplemented by industry averages or migration results of similar users. As data accumulates, the system gradually transitions to a data-driven model, ensuring stable operation across different user lifecycles.

[0019] 3. The dynamic weight fusion mechanism improves recommendation accuracy. By introducing profile similarity scores and data quality scores, the fusion ratio of subjective and objective results is adaptively adjusted, so that the recommendation weights are both individualized and robust.

[0020] 4. Path recommendation under multi-objective optimization: Under the guidance of fusion weights, the path recommendation module performs weighted ranking of cost, time and carbon emissions, and considers realistic conditions such as throughput capacity, time window constraints and scheduling priority, so as to ensure the feasibility and optimality of the recommended solution.

[0021] 5. Iterative optimization and continuous improvement: User selection results and feedback data will flow back to the historical learning module to realize incremental updates and online fine-tuning of the model, continuously improving the matching degree of recommendations and user satisfaction.

[0022] 6. This invention can significantly improve the personalization and intelligence of river-sea intermodal transport route recommendation, and achieve dynamic recommendation effect that is "different for each person and for each order". This makes the recommendation results highly universal and specific, and meets the precise matching needs of high-value users, green transportation companies or time-sensitive goods. Attached Figure Description

[0023] Figure 1 This is a flowchart of the automatic route recommendation algorithm for river-sea intermodal transport in this application; Figure 2 for Figure 1 Flowchart of the automatic recommendation algorithm for a single order; Figure 3 for Figure 1 Flowchart of the algorithm for automatically recommending multiple orders without priority in China; Figure 4 for Figure 1 Flowchart of an algorithm for automatically recommending multiple orders with priority; Figure 5 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0025] A method for recommending river-sea intermodal transport routes based on user preference recognition, such as Figures 1 to 5As shown, it includes: Step S1: Collect user questionnaire results on cost, time and carbon emissions, and generate a scoring vector and an AHP-calculated weight vector respectively; The rating vector is: T01 = ( , , ) in: For cost rating, A score for time. For carbon emissions scoring, T01 is the scoring vector. Questionnaires that directly score cost variables Questionnaires with time variables were scored directly. The questionnaire directly scores the carbon emission variable. The relative score of the cost variable compared to the time variable. The relative score of the time variable compared to the carbon emission variable. The relative scores of cost variables compared to carbon emission variables. , , , , , The pre-configured scoring weights are set to 0.5 by default in this embodiment or can be tuned based on training.

[0026] The process of constructing the weight vector in AHP calculation includes: Step S1-1: Construct the AHP judgment matrix A: Where: 'a' is the importance evaluation value of cost relative to time, and its value is [value missing]. 'b' represents the importance assessment value of cost relative to carbon emissions, and its value is [value missing]. c represents the importance of time relative to carbon emissions, and its value is [value missing]. ; Step S1-2: Calculate the original weights of each objective based on the AHP judgment matrix A as the first and third objective preference weights.

[0027] in: The original weights for the i-th objective are: cost, time, and carbon emissions. Let be the geometric mean of the i-th target. The element in the i-th row and j-th column of the AHP judgment matrix; The system based on the above method includes a user attribute collection module, a preference questionnaire evaluation module, an AHP structure judgment module, an initial weight calculation module, a historical learning and model training module, a dynamic weight fusion module, and a path recommendation module.

[0028] The user attribute acquisition module is responsible for obtaining users' basic attributes, behavioral characteristics, and industry constraints, and storing them in the system as input for subsequent analysis. The preference questionnaire assessment module works in conjunction with the AHP structure judgment module to convert users' subjective preferences for three types of objectives—cost, time, and carbon emissions—into quantitative scores. The initial weight calculation module integrates and normalizes the above scoring results to form an initial preference weight vector; The historical learning and model training module utilizes users' historical order and behavior data to automatically infer and update users' actual preference weights through a multi-task learning model. The dynamic weight fusion module combines subjective preferences with historical learning results to generate the final weight that best represents the user's current needs. The path recommendation module evaluates and ranks multi-objective paths based on this weight, outputs the optimal and alternative solutions, and sends the user's selections and feedback back to the historical learning module to achieve iterative optimization and continuous improvement.

[0029] The user attribute collection module acquires, classifies, encodes, and stores structured feature information of users of the river-sea intermodal transport route recommendation system to support subsequent preference identification, behavior modeling, and personalized recommendations. It includes collection content and a collection process. The structured feature information includes industry category, company size, cargo type, transport frequency, and green compliance requirements. The collection content includes basic user attributes, logistics behavior characteristics, and industry constraint variables. The collection process consists of registration and data entry, backend import, dynamic updates, and data verification.

[0030] Step S2: Fuse the rating vector and the AHP-calculated weight vector to obtain the first and third objective preference weights; Step S2 includes Step S2-1: Calculate the corrected weights for each objective based on their original weights. in: The corrected weight for the i-th objective. This is the subjective rating weighting coefficient, and the recommended initial value is 0.7; Step S2-2: Normalize the corrected weights of each objective to obtain the normalized corrected weights of each objective, which will be used as the preference weights for the second and third objectives. in: , is a temperature coefficient used to control the degree of weight polarization; The output tends to average. The output tends to be extreme; Recommended initial settings Or the system can adaptively set it based on industry experience; The sum of the weights of the three target preferences is 1.

[0031] Step S3: Based on the user's historical order and behavior data, dynamically update the second and third target preference weights in combination with the first and third target preference weights; Step S3 is achieved through the first weight prediction model, which is a Multi-Task Learning policy model based on the XGBoost core. (1) During cold start: Use the output of the preference questionnaire assessment module; (2) During hot start: Integrating questionnaire data with back-derived historical behavior values: 𝜂 is a dynamic weight adjustment parameter that the system can automatically set based on the maturity of user behavior.

[0032] In the structure of the first weighted prediction model, the first K trees form a shared layer, learning common features between tasks. Subsequently, each target is assigned K independent regression trees to learn their individual differences, ultimately outputting three independent weight values. , , ; The first-weighted prediction model uses a multi-task joint loss function: in: For multi-task joint loss function, Let N be the label value of the cost item weight for the i-th task, and N be the number of tasks. The model prediction value for the cost item weights of the i-th task. Let be the label value of the time item weight for the i-th task. The model prediction value for the time term weight of the i-th task. Let the label value be the weight of the carbon emission item in the i-th task. Let be the model-predicted weights of the carbon emission items for the i-th task. This is a complexity regularization term used to prevent overfitting.

[0033] Specifically, the cold start phase: When users first use the system or when there is a lack of sufficient historical order and behavioral data, the system enters a cold start phase. This phase is primarily driven by subjective preference data, specifically: on the one hand, relying on the subjective ratings output by the preference questionnaire assessment module and the AHP structure judgment module, an initial weight vector is directly generated and used as the main basis for decision-making; on the other hand, the historical learning and model training module compensates for the lack of historical data by introducing pre-trained models from industry average weights, similar user profile transfer results, or public sample libraries. The output predicted values ​​are only used as references or correction factors, and their overall weight in recommendation decisions is relatively low. The system also sets dynamically adjusting parameters so that as actual user orders and feedback data gradually accumulate, the reliability of the model predictions gradually increases, smoothly transitioning to the warm start phase.

[0034] The specific steps are as follows: 1) Missing samples were filled using average preference weights based on industry labels; 2) Based on a large historical sample set, XGBoost uses a partially shared structure for pre-training; 3) Limited user behavior → Questionnaire weighting dominates, model predictions serve as reference values; During the warm start phase: Once users have accumulated a certain number of order and transportation selection behaviors in the system, the system enters the warm-up phase. At this point, the system integrates historical behavioral data with subjective questionnaire results and continuously corrects users' implicit preferences through a back-reasoning mechanism. Specifically, the historical learning and model training module uses users' actual route selections, order feedback, and adjustment records as supervisory signals to dynamically update the weights of three target categories: cost, time, and carbon emissions, and achieves incremental training through a multi-task learning model. Simultaneously, the system automatically increases the weight proportion of model predictions based on data volume and behavioral consistency, gradually enhancing its dominance in the final fusion result; while questionnaire weights serve as auxiliary correction items, ensuring that temporary user preferences in individual tasks are captured promptly. Throughout the process, the system supports online fine-tuning and periodic retraining to continuously optimize the model's generalization performance and stability, thereby achieving a smooth transition from initial subjective-driven to data-driven approaches.

[0035] The specific steps are as follows: 1) As orders accumulate, users' actual choices reveal their implicit preferences; 2) Update the label set and add the data to the incremental training; 3) Supports online fine-tuning and periodic model retraining.

[0036] Step S4: Combine the first and third objective preference weights and the second and third objective preference weights to obtain the third and third objective preference weights; In step S4, the third three-objective preference weight is the weighted sum of the first three-objective preference weight and the second three-objective preference weight, where the weight of the second three-objective preference weight is the fusion coefficient γ, and the weight of the first three-objective preference weight is 1-γ. The fusion coefficient γ is: in: To objectively learn the trust coefficient adjustment ratio, To calculate the consistency between user profiles and current model predictions by using structured modeling based on user basic attributes, industry characteristics, historical orders, and questionnaire results, Adjust the data volume / coverage ratio. Scoring the quality of behavioral data The confidence coefficient is based on the prediction.

[0037] In this implementation: α is 0.5, β is 0.4, and γ0 is 0.1.

[0038] (2) Extreme cases are handled as follows: 1) For new users with no prior activity history: Set 𝛾=0 and use the questionnaire results entirely; 2) For high-frequency users with mature profiles and sufficient feedback: 𝛾→1 mainly use model output; 3) The system allows users to manually adjust (±10%), enhancing user engagement; (3) The output is a normalized three-dimensional vector: The final recommended weights must meet the following requirements: The data is passed to the path recommendation module as the basis for the weighted scoring of the three objectives.

[0039] Step S5: Based on the third objective preference weights and the pre-configured overall objective function, calculate the overall objective function value for each candidate intermodal transport route, and rank the candidate intermodal transport routes from best to worst according to their overall objective function values, including: Step S5-1: Construct a directed network containing port nodes, path edges, and transportation modes. Set the following attribute for each edge: path distance For transportation speed Unit transportation cost Unit carbon emission factor throughput per unit time Path set ; Step S5-2: Construct the order structure model and define each order as... ,in: As the starting and ending points, For the type of goods, For transport weight; Let the indicator preference weight vector satisfy... ; Step S5-3: Make basic assumptions and establish constraints for the order structure model to ensure its rationality and solvability. The basic assumptions include: (1) Orders can arrive at different times, and the information is known in advance, which belongs to static or quasi-dynamic scheduling. (2) The loading and unloading capacity and transportation time of ports and routes can be quantitatively estimated and vary with the hour. (3) The costs, emissions, and speeds of each mode of transportation are known or predictable fixed values. (4) The attributes and port service levels of each route segment do not fluctuate due to changes in orders; The constraints include transportation time window constraints, route capacity constraints, scheduling priority constraints, loading and unloading operation time constraints, and route continuity and feasibility constraints.

[0040] (1) The transportation time window constraint is: in: For order O k The actual arrival time For order O k The earliest arrival time, For order O k Latest arrival time; (2) The path capacity constraint is: in: This represents the throughput per unit time of the path edge or port node. This represents the maximum throughput per unit time on the path segment from node i to node j. (3) Scheduling priority constraints (if applicable): If the order priority is The scheduling order satisfies: (4) Loading and unloading time constraints: The total transportation time of the route must include the loading and unloading time. : (5) Path continuity and feasibility constraints: The set of paths should satisfy the requirement of feasible connection from the starting point to the ending point: Step S5-4: Generate multiple candidate intermodal routes for all orders based on order type, where order type includes single order, non-competitive multiple orders, and competitive multiple orders; Step S5-5: Calculate the overall objective function value of each candidate intermodal transport route based on the pre-configured initial overall objective function, and sort the candidate intermodal transport routes from best to worst according to the overall objective function value.

[0041] Consider order sets Each order The quantity of goods is The candidate path set is P; for each path The process of calculating the overall objective function value and constructing the pre-configured initial overall objective function includes: Calculate the cost function: in: For cost function, For order k in path segment e i On the transportation cost, For transportation time; Calculate the time function in: It is a time function. This represents the transportation distance of the route segment. The average speed corresponding to the mode of transportation. These are the loading and unloading times, This refers to the waiting time caused by scheduling and queuing. Calculate the carbon emission function: in: For emission functions, Carbon emissions from the route segment; Construct a pre-configured initial overall objective function: in: , , These are the weights of the cost function, time function, and carbon emission function, respectively, and their sum is 1. The normalized three types of indicators are: in: , , , representing the weighting coefficient, satisfying .

[0042] In step S5-5, based on the pre-configured initial overall objective function, variant objective functions are generated for a single order, multiple orders without competition, and multiple orders with competition, each modeled independently. The modified objective function for a single order considers only one order O. k The transportation process along path p∈P does not involve throughput competition. Calculations are performed using unit time, cost, and carbon emissions. The model is obtained as follows: in: For unit transportation cost, This represents the normalized unit transportation time. Normalized carbon emissions per unit of transportation This represents the minimum unit transportation cost among all candidate route options. This represents the maximum unit transportation cost among all candidate route options. Normalized unit transportation cost The constraints are simplified to: The modified objective function for multiple orders without competition is: in: The total shipping cost for a standardized, non-competitive multi-order shipment. For the standardized total time of multiple orders without competition, The total carbon emissions of multiple orders without competition after standardization.

[0043] The constraint is that the throughput of each order path segment must meet the capacity limit within the selected time period: Preferred, competitive multi-order simulation models are obtained through the following methods: When multiple orders share a path segment within a certain time window ,Right now This will trigger throughput contention, requiring the introduction of a scheduling mechanism; The system is configured with the following scheduling principles: orders are prioritized based on the type of goods ordered. Queue waiting time It will be dynamically changed and incorporated into the indicator calculation; The path simulation order for each order is subject to scheduling constraints; Carbon emissions and costs require additional waiting items: , ; In this case, the scheduling results need to be considered uniformly before evaluating the overall performance. Introducing the performance indicators that await correction after scheduling, the objective function becomes: Among them: the total transportation cost for standardized competitive multi-order shipments is The same applies to standardized time and carbon emissions.

[0044] Scheduling constraints: The automatic recommendation method for a single-order simulation model is as follows: Step Z0411: Initialize order parameters: including goods type c, weight Departure time Read the corresponding unit freight rate and carbon emission factor based on the cargo category; Step Z0412: Traverse candidate paths: Decompose each path into segments and simulate each segment; determine if the job time window is executable; determine if the current throughput meets the transportation demand; if not, introduce waiting time. Dynamically updated time; Step Z0413: Indicator Accumulation and Standardization (1) Cumulative Indicators , , ; (2) Calculate and standardize unit indicators: Step Z0414: Overall Score and Recommendation Output: The path scheme with the lowest recommended score The automatic recommendation method for a multi-order simulation model without competition is as follows: Step Z0421: Enumerate path combinations (1) Generate all the path permutations (Cartesian product) of all orders; (2) Each combination constitutes a scheduling simulation unit; Step Z0422: Simulation Execution Sequence (1) All orders start at the same time, and are scheduled one by one; (2) Perform throughput conflict detection on the path segment: Step Z0423: Dynamic Resource Recovery Mechanism (1) The throughput capacity is reset to zero every day at midnight; (2) The throughput resources are in a shared (non-isolated) state; Step Z0424: Calculation of Combined Comprehensive Score (1) Cumulative combined indicators: , , ; (2) Standardization and weight calculation: Step Z0425: Failed combination removal: If a segment cannot be scheduled within the maximum waiting time, it is marked as a failure and removed.

[0045] For a competitive multi-order simulation model, the automatic recommendation method is as follows: Step Z0431: Set priority strategy: Use product category or order tag as priority parameter; Step Z0432: Sorting the execution order, specifically including: (1) Execute simulations in ascending order of priority, prioritizing orders first; (2) Subsequent orders need to be judged based on the results of the changes in the throughput resources of the former.

[0046] The questionnaire is structured into two categories, comprising six questions, as detailed in the table below: 1. Single-objective tendency scoring (Likelit scale, 1–5 points), as shown in Table 1: Table 1 2. Paired target preference judgment (preference selection, simulated AHP), as shown in Table 2: Table 2 The types of user attributes, behavioral characteristics, rating feedback, and questionnaire characteristics are shown in Table 3: Table 3 The inputs in the dynamic weight fusion module are shown in Table 4: Table 4 If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for recommending river-sea intermodal transport routes based on user preference recognition, characterized in that, include: Step S1: Collect user questionnaire results on cost, time and carbon emissions, and generate a scoring vector and an AHP-calculated weight vector respectively; Step S2: Fuse the rating vector and the AHP-calculated weight vector to obtain the first and third objective preference weights; Step S3: Based on the user's historical order and behavior data, dynamically update the second and third target preference weights in combination with the first and third target preference weights; Step S4: Combine the first and third objective preference weights and the second and third objective preference weights to obtain the third and third objective preference weights; Step S5: Based on the third objective preference weights, calculate the overall objective function value of each candidate intermodal transport route in combination with the pre-configured overall objective function, and sort the candidate intermodal transport routes from best to worst according to the overall objective function value.

2. The method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 1, characterized in that, The scoring vector is: T01=（ , , ）, , , , in: For cost rating, A score for time. For carbon emissions scoring, T01 is the scoring vector. Questionnaires that directly score cost variables Questionnaires with time variables were scored directly. The questionnaire directly scores the carbon emission variable. The relative score of the cost variable compared to the time variable. The relative score of the time variable compared to the carbon emission variable. The relative scores of cost variables compared to carbon emission variables. , , , , , For pre-configured scoring weights.

3. The method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 2, characterized in that, The process of constructing the weight vector for AHP calculation includes: Step S1-1: Construct the AHP judgment matrix A: , Where: 'a' is the importance evaluation value of cost relative to time, and its value is [value missing]. 'b' represents the importance assessment value of cost relative to carbon emissions, and its value is [value missing]. c represents the importance of time relative to carbon emissions, and its value is [value missing]. ; Step S1-2: Calculate the original weights of each objective based on the AHP judgment matrix A as the first and third objective preference weights; , , in: The original weights for the i-th objective are: cost, time, and carbon emissions. Let be the geometric mean of the i-th target. Let be the element in the i-th row and j-th column of the AHP judgment matrix.

4. The method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 3, characterized in that, Step S2 includes Step S2-1: Calculate the corrected weights for each objective based on their original weights. , in: The corrected weight for the i-th objective. This refers to the subjective rating weighting coefficient. Step S2-2: Normalize the corrected weights of each objective to obtain the normalized corrected weights of each objective as the preference weights of the second and third objectives.

5. The method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 1, characterized in that, Step S3 is implemented through a first weight prediction model, which is a Multi-Task Learning policy model based on the XGBoost core. In the structure of the first weighted prediction model, the first K trees form a shared layer, learning common features between tasks. Subsequently, each target is assigned K independent regression trees to learn their respective differences, ultimately outputting three independent weight values. , , ; The first weighted prediction model uses a multi-task joint loss function: , in: For multi-task joint loss function, Let N be the label value of the cost item weight for the i-th task, and N be the number of tasks. The model prediction value for the cost item weights of the i-th task. Let be the label value of the time item weight for the i-th task. The model prediction value for the time term weight of the i-th task. Let the label value be the weight of the carbon emission item in the i-th task. Let be the model-predicted weights of the carbon emission items for the i-th task. This is a complexity regularization term used to prevent overfitting.

6. The method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 1, characterized in that, In step S4, the third three-target preference weight is the weighted sum of the first three-target preference weight and the second three-target preference weight, wherein the weight of the second three-target preference weight is the fusion coefficient γ, and the weight of the first three-target preference weight is 1-γ. The fusion coefficient γ is: , in: To objectively learn the trust coefficient adjustment ratio, To calculate the consistency between user profiles and current model predictions by using structured modeling based on user basic attributes, industry characteristics, historical orders, and questionnaire results, Adjust the data volume / coverage ratio. Scoring the quality of behavioral data The confidence coefficient is based on the prediction.

7. The method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 1, characterized in that, Step S5 includes: Step S5-1: Construct a directed network containing port nodes, path edges, and transportation modes. Set the following attribute for each edge: path distance For transportation speed Unit transportation cost Unit carbon emission factor throughput per unit time Path set ; Step S5-2: Construct the order structure model and define each order as... ,in: As the starting and ending points, For the type of goods, For transport weight; Let the indicator preference weight vector satisfy... ; Step S5-3: Make basic assumptions and establish constraints for the order structure model to ensure the model's rationality and solvability. The basic assumptions include: (1) Orders can arrive at different times, and the information is known in advance, which belongs to static or quasi-dynamic scheduling. (2) The loading and unloading capacity and transportation time of ports and routes can be quantitatively estimated and vary with the hour. (3) The costs, emissions, and speeds of each mode of transportation are known or predictable fixed values. (4) The attributes and port service levels of each route segment do not fluctuate due to changes in orders; The constraints include transportation time window constraints, route capacity constraints, scheduling priority constraints, loading and unloading operation time constraints, and route continuity and feasibility constraints. The transportation time window constraint is as follows: , in: For order O k The actual arrival time For order O k The earliest arrival time, For order O k Latest arrival time; The path capacity constraint is: , in: This represents the throughput per unit time of the path edge or port node. This represents the maximum throughput per unit time on the path segment from node i to node j. Step S5-4: Generate multiple candidate intermodal routes for all orders based on order type, wherein the order type includes single order, non-competitive multiple orders, and competitive multiple orders; Step S5-5: Calculate the overall objective function value of each candidate intermodal transport route based on the pre-configured initial overall objective function, and sort the candidate intermodal transport routes from best to worst according to the overall objective function value.

8. The method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 7, characterized in that, Consider order sets Each order The quantity of goods is The candidate path set is P; for each path The process of calculating the overall objective function value and constructing the pre-configured initial overall objective function includes: Calculate the cost function: , in: For cost function, For order k in path segment e i On the transportation cost, For transportation time; Calculate the time function , in: It is a time function. This represents the transportation distance of the route segment. The average speed corresponding to the mode of transportation. These are the loading and unloading times, This refers to the waiting time caused by scheduling and queuing. Calculate the carbon emission function: , in: For emission functions, Carbon emissions from the route segment; Construct a pre-configured initial overall objective function: , in: , , These are the weights of the cost function, time function, and carbon emission function, respectively, and the sum of the three is 1.

9. A method for recommending river-sea intermodal transport routes based on user preference recognition according to claim 7, characterized in that, In step S5-5, modified objective functions are generated based on the pre-configured initial overall objective function, for single orders, multiple orders without competition, and multiple orders with competition, which are modeled independently. The modified objective function for a single order considers only one order O. k The transportation process along path p∈P does not involve throughput competition. Calculations are performed using unit time, cost, and carbon emissions. The model is obtained as follows: , , in: For unit transportation cost, This represents the normalized unit transportation time. Normalized carbon emissions per unit of transportation This represents the minimum unit transportation cost among all candidate route options. This represents the maximum unit transportation cost among all candidate route options. Normalized unit transportation cost The modified objective function for multiple orders without competition is: , in: The total shipping cost for a standardized, non-competitive multi-order shipment. For the standardized total time of multiple orders without competition, The total carbon emissions of multiple orders without competition after standardization.

10. A river-sea intermodal transport route recommendation device based on user preference recognition, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-9.

Images