MaaS package fine design and personalized recommendation system based on large model

Abstract

The invention discloses a MaaS package refined design and personalized recommendation system based on a large model, and the system comprises a data end which is used for facing a MaaS platform operator and providing a multi-source spatio-temporal data fusion service; the enterprise terminal provides a MaaS package fine design service, and an enterprise can analyze a user portrait and a behavior mode thereof through an enterprise terminal Chat large model, so that MaaS packages for different users and different travel activity modes are intelligently generated; the user side is used for providing MaaS package personalized recommendation service for individual travelers, dividing users into commuting, business and tourist crowds, matching optimal MaaS packages for the crowds according to historical travel activity modes of the crowds, and providing basic travel service; the basic travel service comprises destination search, multi-mode travel planning and travel navigation. According to the invention, an integrated travel service is provided, a refined MaaS package is designed, and a personalized MaaS package is recommended.

Description

Technical Field

[0001] The present invention relates to the technical field of travel software systems, and particularly relates to a refined design and personalized recommendation system for MaaS packages based on large models. Background Art

[0002] "Mobility as a Service (MaaS)" refers to the integration of travel services in multiple modes, such as regular buses, rail transit, taxis, online car-hailing (shared or solo rides), car-sharing, customized buses, long-term car rentals, and shared bicycles, through a mobile application interface to provide users with personalized travel packages.

[0003] According to the development maturity and service integration level of MaaS, the industry generally divides MaaS into 5 levels. MaaS at level L0 is non-integrated, and various transportation modes operate independently. Among them, shared travel service platforms are representative, such as Uber and Didi Chuxing. MaaS at level L1 realizes the integration of travel information and provides users with services such as multi-modal travel planning and price queries. Traditional map software services are representative, such as Google Map and AutoNavi Map. MaaS at level L2 realizes the integration of reservation and payment. On the basis of multi-modal travel planning, it realizes the reservation and payment of intermodal travel services, such as Hannover in Germany and Suishenxing (one-code access) in Shanghai. MaaS at level L3 is the integration of travel service supply, which packages multiple transportation services into customized travel packages. Such packages stipulate the number of times, time, mileage, or price that users can use various transportation modes within a month (or a quarter) within a certain area. Users only need to pay the package fee once, instead of paying separately for the combined transportation modes. For example, UbiGo in Gothenburg, Sweden, and Whim in Helsinki, Finland. MaaS at level L4 is the integration of social goals, not only limited to meeting the travel needs of users, but also embedding social values into the price of MaaS packages, guiding the transformation of user behavior through price signals. The MaaS platform encourages people to use urban public transportation, reduce private car travel, relieve road congestion, and ultimately reduce carbon emissions, making the city cleaner and more livable through the implementation of behavior regulation policies and incentive measures. However, currently no MaaS products have reached this level.

[0004] It can be seen that most domestic MaaS platforms are currently at levels L1 and L2, only realizing the planning, reservation, and payment functions of a single travel mode. Some foreign MaaS products have reached level L3 and can package multiple transportation services into travel packages. However, the types of packages are still relatively single at present. Users can only choose from a limited number of options and cannot design personalized MaaS packages according to the differentiated travel needs of users, making it difficult to effectively attract users to subscribe to MaaS packages. Therefore, in order to achieve the development vision of MaaS, it is urgent to accelerate the development and construction of MaaS at levels L3 and L4. Summary of the Invention

[0005] Aiming at the deficiencies in the existing technology, the purpose of the present invention is to provide a refined design and personalized recommendation system for MaaS packages based on large models, which provides an integrated travel service, designs refined MaaS packages, and recommends personalized MaaS packages. To achieve the above object and other advantages of the present invention, there is provided a refined design and personalized recommendation system for MaaS packages based on large models, including:

[0006] A data terminal, which is used to provide multi-source spatio-temporal data fusion services for MaaS platform operators;

[0007] An enterprise terminal, which is used to provide refined design services for MaaS packages for travel service providers. Enterprises can analyze user portraits and their behavior patterns through the enterprise terminal Chat large model, so as to intelligently generate MaaS packages for different users and different travel activity patterns;

[0008] A user terminal, which is used to provide personalized recommendation services for MaaS packages for individual travelers, divides users into commuting, business, and tourism groups, matches the best MaaS packages according to their historical travel activity patterns, and provides basic travel services. The basic travel services include destination search, multi-modal travel planning, and travel navigation;

[0009] Among them, the enterprise terminal includes a multi-day travel activity pattern recognition module and a refined design module for MaaS packages;

[0010] The multi-day travel activity pattern recognition module is used for extracting multi-day travel activity chains and recognizing travel activity patterns;

[0011] The refined design module for MaaS packages is used to refine the design of corresponding MaaS packages for dozens of different travel activity patterns, including the travel mode combination of each package, the number of times or mileage of freely using each travel mode within a specified period and area, and value-added services.

[0012] Preferably, the client includes a personalized recommendation module for MaaS packages and a basic travel service module for MaaS. The personalized recommendation module for MaaS packages includes a user profiling unit, which is used to initially classify users into three categories: commuting, business, and tourism. Based on features such as the user's historical travel habits, preferences, and self-selected social attribute tags, a profile is generated that comprehensively represents the spatio-temporal heterogeneity of each user.

[0013] Preferably, the personalized recommendation module for MaaS packages is used to perform similarity matching between the spatio-temporal profile of the user and all MaaS packages provided by the enterprise side, so as to accurately recommend personalized MaaS packages for users of different groups. Users can decide whether to select or adjust the package according to their own needs.

[0014] Preferably, the basic travel service module for MaaS includes a package balance query unit, a destination search, a multi-modal travel planning unit, a one-code pass unit, a travel navigation unit, and a green travel points unit.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] (1) Provide an integrated travel service: Based on the MaaS platform, integrate various transportation modes such as carsharing, online car-hailing, taxis, regular buses, rail transit, shared (electric) bicycles, etc., so as to package multiple transportation services into an integrated travel package and provide it to users, while realizing one-code pass.

[0017] (2) Design refined MaaS packages: Based on the fusion of multi-source spatio-temporal data of MaaS, extract the spatio-temporal semantic information of multi-day travel activity chains, and use large models to perform representation learning on travel activity chains. Identify diverse travel activity patterns of users through semantic clustering, and thus design refined MaaS packages according to the behavior patterns of different users.

[0018] (3) Recommend personalized MaaS packages: Integrate multi-dimensional spatio-temporal feature vectors such as user behavior patterns, travel preferences, and individual characteristics, construct a user profile considering spatio-temporal context, and at the same time convert the content items in all MaaS packages into vector representations, so as to perform multi-dimensional matching between users and MaaS packages according to semantic similarity, thereby realizing the personalized recommendation of "one size fits one person" for MaaS packages. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the architecture of a system for refined design and personalized recommendation of MaaS packages based on a large model according to the present invention.

[0020] Figure 2It is a technical schematic diagram of a refined design and personalized recommendation system for MaaS packages based on large models according to the present invention.

[0021] Figure 3 It is a technical schematic diagram of personalized recommendation for MaaS packages of user portraits of a refined design and personalized recommendation system for MaaS packages based on large models according to the present invention. Specific implementation manners

[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] Refer to Figure 1 , a refined design and personalized recommendation system for MaaS packages based on large models, includes: a data terminal, which is used to provide a multi-source spatio-temporal data fusion service for the MaaS platform operator; the multi-source spatio-temporal data fusion service includes user attribute data, travel order data, GPS trajectory data, social media data, application interaction data, and POI data. The data terminal is used for multi-source data collection and storage. Specifically, based on the MaaS platform, a data opening and sharing mechanism between travel service providers and MaaS operators is established to collect dynamic spatio-temporal data of various transportation modes such as car-sharing, online car-hailing, taxis, regular buses, rail transit, shared bicycles, and shared electric bicycles, including orders and GPS trajectories for multiple days, and record the interaction data generated by users on the platform. At the same time, static facility data of multi-modal transportation networks, station hubs, and points of interest are collected. Finally, distributed storage of massive data is realized based on Hadoop HDFS, and load balancing and high availability of data access are achieved.

[0024] The data terminal is used for multi-modal transportation data fusion processing: fusing the multi-modal transportation travel spatio-temporal data collected by the MaaS platform, preprocessing the multi-source heterogeneous big data based on the data warehouse ETL algorithm, including data cleaning, data integration, data reduction, and data transformation, and establishing a standardized expression system with different parameters to achieve semantic unity of multi-modal transportation data.

[0025] The data terminal is used for spatio-temporal big data computing and visualization: Based on in-memory computing of Spark, adopting the elastic distributed dataset technology, after importing the computing data into the memory of the computing nodes, it realizes the distributed high-speed transformation and processing of massive data; using spatio-temporal visualization technologies such as GIS, D3.js, Leaflet, and Kepler.gl to transform complex multi-source spatio-temporal data into intuitive graphs and maps to better understand the spatio-temporal distribution and dynamic changes in the data.

[0026] Enterprise terminal, the enterprise terminal is used to provide refined design services for MaaS packages for travel service providers. Enterprises can analyze user portraits and their behavior patterns through the enterprise terminal Chat large model, and thus intelligently generate MaaS packages for different users and different travel activity patterns.

[0027] Among them, the enterprise terminal includes a multi-day travel activity pattern recognition module and a refined design module for MaaS packages.

[0028] The multi-day travel activity pattern recognition module is used for multi-day travel activity chain extraction and travel activity pattern recognition; the multi-day travel activity chain extraction is specifically that the travel service provider can extract multi-day travel activity chain information of a specified area, specified time range, and specified user group from the data terminal of the MaaS platform through dialogue interaction with the enterprise terminal Chat large model, including spatio-temporal features such as user ID, travel segment ID, travel time, travel space, travel distance, travel mode, and travel purpose.

[0029] The travel activity pattern recognition is specifically that the travel service provider can conduct spatio-temporal feature analysis and visualization on the user's multi-day travel activity chain through dialogue interaction with the enterprise terminal Chat large model, and represent it as spatio-temporal semantics. According to the semantic similarity, all user travel activity chains are clustered and grouped, and each group represents a travel activity pattern with similar spatio-temporal features. By analyzing the continuity, periodicity, and repeatability laws of each pattern, typical travel activity patterns are identified.

[0030] The refined design module for MaaS packages: The travel service provider can, through dialogue interaction with the enterprise terminal Chat large model, refine the design of corresponding MaaS packages for dozens of different travel activity patterns. The main contents include the travel mode combination of each package, the number of times or mileage of freely using each travel mode within a specified period, such as daily, weekly, monthly, and within a region, as well as value-added services, etc., to meet the differentiated needs of different user groups. For example Figure 1 In Package 1: It is stipulated that the user is allowed to use the bus 15 times + the subway 25 times + shared bicycles for 30 kilometers within a week.

[0031] The refined design module of the MaaS package first integrates multi-source spatio-temporal data based on the MaaS platform to extract the spatio-temporal semantic information of the complete multi-day travel activity chain of users. Then, based on the spatio-temporal semantic embedding model, the travel activity chain is represented as a semantic vector, and different travel activity patterns are identified through semantic clustering. For different types of behavior patterns, refined travel packages are generated. As Figure 2 shown, the specific operation steps are as follows:

[0032] S1: Deduction of multi-day travel activity chain based on spatio-temporal association;

[0033] According to the user ID, order time, and spatial location, the spatio-temporal association analysis method is used to extract the travel segments and their spatio-temporal characteristics of all transportation modes of each user within multiple days from different transportation data sources. At the same time, combined with the travel log data of volunteers and POI data, the travel purpose of each travel segment, that is, the activity type, is inferred, and then connected in chronological order, so as to completely deduce the whole process information of the user's travel activity chain and restore the spatio-temporal panorama of the user's travel activity chain.

[0034] S2: Construction of spatio-temporal semantic knowledge graph of travel activity chain;

[0035] Express the spatio-temporal characteristics of the user's travel activity chain in natural language, including semantic information such as user ID, travel segment ID, travel time, travel space, travel distance, travel mode, travel purpose, etc., and add spatio-temporal tags, such as "morning rush hour", "suburban area", etc. Then extract entities, attributes, attribute values, and their mutual relationships from them, form an ontological knowledge expression on this basis, and then convert it into a graph structure for storage through knowledge mapping to construct a spatio-temporal semantic knowledge graph of the travel activity chain.

[0036] S3: Spatio-temporal semantic embedding of travel activity chain based on large model;

[0037] First, convert the spatio-temporal semantic information of the user's multi-day travel activity chain into text form. Each travel activity feature is encoded as a "word", and travel activity features such as travel time, travel mode, and travel distance. The single-day travel chain composed of these travel activity features is encoded as a "sentence", and the multi-day travel activity chain is encoded as a "paragraph". Then input it into the BERT large model, and at the same time load the spatio-temporal semantic knowledge graph of the travel activity chain. Optimize and fine-tune the pre-trained BERT large model by integrating domain knowledge to construct a spatio-temporal semantic embedding model, and represent the spatio-temporal characteristics of each user's travel activity chain as a comprehensive semantic vector to identify the travel activity patterns of users in the vector space.

[0038] S4: Recognition of user travel activity patterns based on semantic clustering;

[0039] Extract the semantic vector representations of the spatio-temporal characteristics from the user travel activity chains of different periods, such as 1 day, 3 days, and 7 days. Measure the semantic similarity between vectors by calculating the cosine similarity. Then apply clustering algorithms such as AP, DBSCAN, spectral clustering, and sparse subspace clustering (SSC) to group the semantically similar vectors, and use internal evaluation metrics such as DBI and VRC to evaluate the quality of clustering and optimize the clustering results. Finally, visualize the clustering results, analyze the spatio-temporal patterns therein, and attempt to understand the travel activity patterns represented by each cluster.

[0040] S5: Design of MaaS packages based on travel activity patterns;

[0041] Calculate the average word frequencies and weights of spatio-temporal semantics such as different travel times, travel regions, travel modes, and travel distances in typical travel activity patterns based on text analysis algorithms. Then construct a spatio-temporal semantic matrix of travel activities to depict the differentiated behavior patterns of users. Furthermore, design refined MaaS packages for different user travel activity patterns. The main contents include the combination of travel modes for each package and the number of times or mileage of freely using each travel mode within a specified time and region.

[0042] User side: The user side is used to provide personalized recommendation services for MaaS packages to individual travelers, classify users into commuting, business, and tourism groups, match the best MaaS packages for them according to their historical travel activity patterns, and provide basic travel services, including destination search, multimodal travel planning, and travel navigation.

[0043] The user side includes a personalized recommendation module for MaaS packages and a basic travel service module for MaaS. The personalized recommendation module for MaaS packages includes a user portrait unit, which is used to initially classify users into three categories: commuting, business, and tourism. According to characteristics such as the historical travel habits, preferences of users, and the social attribute tags they choose, a portrait is generated that comprehensively represents the spatio-temporal heterogeneity of each user.

[0044] Furthermore, the personalized recommendation module for MaaS packages is used to perform similarity matching with all MaaS packages provided by the enterprise side based on the spatio-temporal portraits of users, so as to accurately recommend personalized MaaS packages for different groups of users. Users can decide whether to select or adjust the packages according to their own needs. For example: recommend a travel package mainly based on bus + subway for commuting people; recommend a travel package mainly based on car rental and hotel value-added services for business people; recommend a travel package mainly based on online car-hailing + shared bicycle and value-added services for scenic spot tickets for tourism people.

[0045] The MaaS package personalized recommendation module integrates multi-dimensional spatiotemporal feature vectors such as user behavior patterns, travel preferences, and individual characteristics to construct a user profile that considers spatiotemporal contexts, so as to achieve more accurate and personalized MaaS package recommendations, such as Figure 3 As shown, the specific process includes:

[0046] S1: User spatiotemporal preference extraction based on topic model;

[0047] The user's spatiotemporal preference labels, such as travel time, travel area, travel mode, travel purpose and travel distance, are extracted from the spatiotemporal semantics of the user's multi-day travel activity chain, and a behavioral semantic document is constructed for each user. The Latent Dirichlet Allocation (LDA) topic model is then used to train these behavioral semantic document sets to learn the document-topic distribution and topic-word distribution. The user's travel spatiotemporal preference vector is constructed by identifying the implicit topics in the document and calculating the probability distribution of each document corresponding to these topics.

[0048] S2: Construction of crowd portrait considering users’ spatial and temporal heterogeneity;

[0049] User attribute labels (including user gender, age, region, education level, date of birth, occupation, etc.) are extracted from the user registration data of the MaaS platform and merged with the previously generated user travel time and space preference vectors and travel activity pattern vectors to generate a portrait vector that comprehensively reflects the user's time and space heterogeneity. The cosine similarity is used to calculate the distance between user portrait vectors, and then clustering algorithms (such as K-mean, hierarchical clustering, etc.) are applied to divide users into several groups. Users within each group have similar attributes and behavioral preferences, and then a population portrait that can describe their common characteristics is constructed for each segmented group.

[0050] S3: Personalized recommendation of MaaS packages based on semantic similarity;

[0051] First, based on the same semantic embedding model, all content items in the MaaS packages are converted into vector representations, and users and MaaS packages are matched in multiple dimensions based on semantic similarity. Then, a content-based personalized recommendation algorithm is designed to calculate the similarity between each user portrait vector and the travel package vector, and recommend the N most similar MaaS packages to the user. At the same time, a recommendation algorithm based on collaborative filtering is designed, which takes users who have already selected travel packages as the target, searches for the N most similar users from the same group as neighbors, and recommends the same MaaS packages to these N users.

[0052] S4: Optimization of MaaS package recommendation strategy based on reinforcement learning;

[0053] Through the MaaS platform, users’ selection behaviors and feedback on recommended packages are collected online. A deep reinforcement learning algorithm is used to continuously update user portraits during user interaction and optimize recommendation strategies to ensure the real-time and accuracy of recommended content, thereby improving user experience and satisfaction.

[0054] Furthermore, the MaaS basic travel service module includes a package balance query unit, a destination search, a multi-modal travel planning unit, a one-code pass unit, a travel navigation unit, and a green travel points unit. When a user uses green travel, the platform will calculate the corresponding carbon emission reduction according to the official carbon inclusive methodology and convert it into carbon points. These points can be exchanged for goods or rights in the Carbon Inclusive Mall. When a user chooses a MaaS package, the user can query the destination and plan a multi-modal travel plan according to actual travel needs, and use the travel QR code to achieve multi-modal one-stop travel. When the trip is completed, the used travel services will be automatically deducted from the package quota, and the user can also check the balance of the MaaS package and historical travel orders at any time.

[0055] The number of devices and processing scales described here are used to simplify the description of the present invention, and the application, modification and variation of the present invention are obvious to those skilled in the art. Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation mode, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, other modifications can be easily realized, so without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to the specific details and the legends shown and described here.

Claims

1. A MaaS package refined design and personalized recommendation system based on a large model, characterized by: include: A data terminal, which is used to provide multi-source spatiotemporal data fusion services to MaaS platform operators; Enterprise side: The enterprise side is used to provide MaaS package refined design services to travel service providers. Enterprises can use the enterprise side Chat big model to analyze user portraits and their behavior patterns, thereby intelligently generating MaaS packages for different users and different travel activity patterns; The user terminal is used to provide personalized MaaS package recommendation services for individual travelers, classify users into commuting, business and tourism groups, and match them with the best MaaS packages based on their historical travel activity patterns, as well as provide basic travel services, including destination search, multi-modal travel planning, and travel navigation; The enterprise side includes a multi-day travel activity pattern recognition module and a MaaS package refinement design module; The multi-day travel activity pattern recognition module is used for multi-day travel activity chain extraction and travel activity pattern recognition; The MaaS package refined design module is used to refine the design of corresponding MaaS packages for dozens of different travel activity modes, including the travel mode combination of each package, the number of times or mileage of free use of each travel mode within a specified period and area, and value-added services.

2. The MaaS package refined design and personalized recommendation system based on a large model as claimed in claim 1, characterized in that: The user end includes a MaaS package personalized recommendation module and a MaaS basic travel service module. The MaaS package personalized recommendation module includes a user portrait unit. The user portrait unit is used to preliminarily divide users into three categories: commuting, business, and tourism. According to the user's historical travel habits, preferences, and social attribute labels selected by the user, a portrait that comprehensively represents the temporal and spatial heterogeneity of each user is generated.

3. A MaaS package refined design and personalized recommendation system based on a large model as described in claim 2, characterized in that: The MaaS package personalized recommendation module is used to perform similarity matching with all MaaS packages provided by the enterprise side based on the user's spatiotemporal portrait, so as to accurately recommend personalized MaaS packages to users of different groups. Users can decide whether to select or adjust the package according to their own needs.

4. The MaaS package refined design and personalized recommendation system based on a large model as described in claim 2, characterized in that: The MaaS basic travel service module includes package balance query unit, destination search, multi-modal travel planning unit, one-code access unit, travel navigation unit and green travel points unit.

5. The MaaS package refined design and personalized recommendation system based on a large model as claimed in claim 1, characterized in that: The specific operation of the MaaS package refined design module includes the following steps: S1. Multi-day travel activity chain deduction based on spatiotemporal correlation; S2, construction of spatiotemporal semantic knowledge graph of travel activity chain; S3, spatiotemporal semantic embedding of travel activity chains based on large models; S4, user travel activity pattern recognition based on semantic clustering; S5. Design of MaaS packages based on travel activity patterns.

6. A MaaS package refined design and personalized recommendation system based on a large model as claimed in claim 2, characterized in that: The specific operation of the MaaS package personalized recommendation module includes the following steps: S1. User spatiotemporal preference extraction based on topic model; S2, constructing crowd portraits considering users’ spatial and temporal heterogeneity; S3, personalized recommendation of MaaS packages based on semantic similarity; S4. Optimization of MaaS package recommendation strategy based on reinforcement learning.

7. The MaaS package refined design and personalized recommendation system based on a large model as claimed in claim 5, characterized in that: Specifically, step S1 uses a spatiotemporal correlation analysis method to extract the travel segments and spatiotemporal characteristics of all transportation modes of each user over multiple days from different transportation data sources based on the user ID, order time, and spatial location. At the same time, the travel log data of volunteers and POI data are combined to infer the travel purpose of each travel segment, and they are connected in chronological order, thereby completely deducing the full process information of the user's travel activity chain and restoring the spatiotemporal panorama of the user's travel activity chain.

8. The MaaS package refined design and personalized recommendation system based on a large model as claimed in claim 5, characterized in that: The step S2 specifically expresses the spatiotemporal characteristics of the user's travel activity chain in natural language, and adds spatiotemporal labels, extracts entities, attributes, attribute values ​​and their relationships, forms an ontological knowledge expression, and then converts it into a graph structure storage through knowledge mapping to construct a spatiotemporal semantic knowledge graph of the travel activity chain.

9. The MaaS package refined design and personalized recommendation system based on a large model as claimed in claim 5, characterized in that: The step S3 specifically converts the spatiotemporal semantic information of the user's multi-day travel activity chain into text form, where each travel activity feature is encoded as a "word", a single-day travel chain composed of travel activity features is encoded as a "sentence", and a multi-day travel activity chain is encoded as a "paragraph", and then input into the BERT large model, and at the same time loads the spatiotemporal semantic knowledge graph of the travel activity chain. By integrating domain knowledge, the pre-trained BERT large model is optimized and fine-tuned to construct a spatiotemporal semantic embedding model, and the spatiotemporal features of each user's travel activity chain are represented as a comprehensive semantic vector, so as to identify the user's travel activity pattern in the vector space.

10. The MaaS package refined design and personalized recommendation system based on a large model as claimed in claim 5, characterized in that: The step S4 specifically extracts the semantic vector representation of the spatiotemporal characteristics from the user travel activity chains of different periods, measures the semantic similarity between vectors by calculating the cosine similarity, groups the semantically similar vectors by a clustering algorithm, evaluates the quality of clustering by using DBI and VRC internal evaluation indicators, optimizes the clustering results, and finally visualizes the clustering results, analyzes the spatiotemporal laws therein, and attempts to understand the travel activity pattern represented by each cluster.

Images