A travel recommendation method, device and vehicle

Abstract

The application relates to a travel recommendation method and device and a vehicle, and relates to the technical field of vehicles, and aims to improve the on-the-way degree and fitting accuracy of a travel recommendation result and a current driving scene, and comprises the following steps: determining a candidate interest point set according to a travel voice text of a user; generating a historical access interest point set based on historical access interest points of the user; determining a space-time relationship matrix based on the candidate interest point set and the historical access interest point set; and recommending a travel interest point for the user based on the space-time relationship matrix, the historical access interest point set and the candidate interest point set.

Description

Technical Field

[0001] This invention relates to the field of vehicle technology, specifically to a travel recommendation method, device, and vehicle. Background Technology

[0002] With the rapid iteration of intelligent connected vehicle technology, the smart cockpit has gradually evolved into a "third living space" for users. Among the many services offered by smart cockpits, location-based services and Point of Interest (POI) recommendations are high-frequency and core human-computer interaction scenarios. However, most POI recommendation methods are designed for mobile internet devices and are ill-suited to the dynamic scenarios of smart cockpits.

[0003] Therefore, existing POI recommendation methods still face the challenge of providing intelligent recommendations that are highly relevant and accurately match the current driving scenario. Summary of the Invention

[0004] This invention provides a travel recommendation method, device, and vehicle, aiming to improve the accuracy and relevance of travel recommendation results to the current driving scenario.

[0005] Firstly, this application provides a travel recommendation method, comprising: determining a candidate point of interest set based on the user's travel voice text; generating a historical point of interest set based on the user's historical access points of interest; wherein the historical point of interest in the historical access point of interest set includes access time features and cabin status features; determining a spatiotemporal relationship matrix based on the candidate point of interest set and the historical point of interest set; wherein the spatiotemporal relationship matrix is ​​used to characterize the road network detour cost between a first target node and a second target node; the first target node is a node in the candidate point of interest set; the second target node is a node in the historical point of interest set; and recommending travel points of interest to the user based on the spatiotemporal relationship matrix, the historical point of interest set, and the candidate point of interest set.

[0006] Based on the aforementioned technical means, this application determines a set of candidate points of interest (POIs) using a large language model based on the user's travel voice text. It leverages the semantic understanding capabilities of the large language model to comprehend the user's intent and extract POIs relevant to travel needs, reducing interference from irrelevant POIs in subsequent processing. A set of historical access POIs is generated based on the user's historical access POIs, and these historical POIs include access time features and cabin status features, enabling each historical POI to carry historical behavioral information in both the access time and cabin status dimensions. A spatiotemporal relationship matrix is ​​determined based on the candidate POI set and the historical POI set, and this matrix is ​​used to characterize the relationship between the first target node and the second target node. The network detour cost is calculated by considering two target nodes: the first target node is a node in the candidate point of interest set, and the second target node is a node in the historical visited point of interest set. This ensures that the spatiotemporal relationship matrix contains the actual network travel cost after adding the node to the network. Based on the spatiotemporal relationship matrix, the historical visited point of interest set, and the candidate point of interest set, the system recommends travel points of interest to users. During the recommendation process, it integrates the long-term behavioral patterns reflected by the user's historical access time characteristics and cabin status characteristics, as well as the network detour cost constraints between candidate points of interest and historical visited points of interest. This ensures that the recommendation results not only conform to the user's long-term travel habits but also meet the actual accessibility of the current network, achieving a highly convenient and accurate recommendation of travel points of interest that matches the current driving scenario.

[0007] Furthermore, a first feature matrix corresponding to the set of historical access points of interest is determined; based on the spatiotemporal relationship matrix, attention learning is performed on the first feature matrix to obtain a second feature matrix. The attention of the target historical access points of interest in the first feature matrix is ​​determined based on the road network detour cost between the target historical access points of interest and the target candidate points of interest; the target candidate points of interest are candidate points of interest in the spatiotemporal relationship matrix that have a road network detour relationship with the target historical access points of interest; based on the set of candidate points of interest and the second feature matrix, the target points of interest are determined; and the target points of interest are recommended to the user.

[0008] Based on the above technical means, attention learning is performed on the first feature matrix based on the spatiotemporal relationship matrix, so that the attention weight of the target historical visit interest point in the first feature matrix is ​​determined based on the road network detour cost between the target historical visit interest point and the target candidate interest point. The target candidate interest point is the candidate interest point in the spatiotemporal relationship matrix that has a road network detour relationship with the target historical visit interest point. Thus, the detour time cost and detour spatial distance cost in the actual road network are introduced as direct constraints on the attention weight into the feature learning process, so that each historical visit interest point feature in the second feature matrix carries the road network detour cost information between it and the candidate interest point.

[0009] Furthermore, based on the travel voice text, a travel intention vector is determined; based on the travel intention vector, attention learning is performed on the second feature matrix to determine the third feature matrix; based on the candidate interest point set and the third feature matrix, the target interest point is determined.

[0010] Based on the aforementioned technical means, by determining the travel intention vector based on travel voice text, and performing attention learning on the second feature matrix based on the travel intention vector, the attention weights of historical visit interest points in the third feature matrix are constrained by the current travel intention. This integrates the user's current voice expression needs and preferences into the feature aggregation process of historical visit interest points, so that each feature of historical visit interest points in the third feature matrix is ​​selectively enhanced or suppressed according to its correlation with the current travel intention, thereby achieving the fusion of the user's long-term behavioral patterns and current real-time needs at the feature level.

[0011] Furthermore, based on the third feature matrix, the travel score of each candidate point of interest in the candidate point of interest set is determined; the number of candidate points of interest with the highest travel scores are then selected as the target points of interest.

[0012] Based on the above technical means, the travel score of each candidate point of interest in the candidate point of interest set is determined by the third feature matrix, and a preset number of candidate points of interest with the highest travel scores are determined as target points of interest. This makes the selection and ranking of candidate points of interest have a quantitative score basis. The travel score integrates the comprehensive feature information after learning the dual attention of historical visit points of interest through road network detour cost constraints and travel intention vectors, so that the preset number of candidate points of interest with the highest ranking simultaneously meet the road network route conformity constraints and the current travel intention matching.

[0013] Furthermore, the query matrix, key matrix, and value matrix corresponding to the first feature matrix are determined; the scaled dot product attention matrix of the query matrix and key matrix is ​​determined; the attention weight matrix is ​​determined based on the fusion matrix of the scaled dot product attention matrix and the fourth feature matrix; wherein the fourth feature matrix is ​​the product of the first feature matrix and the scaling scalar; and the second feature matrix is ​​determined based on the attention weight matrix and the value matrix.

[0014] Based on the aforementioned technical means, by determining the query matrix, key matrix, and value matrix corresponding to the first feature matrix, and determining the scaled dot product attention matrix of the query matrix and key matrix, the attention weight matrix is ​​determined based on the fusion matrix of the scaled dot product attention matrix and the fourth feature matrix. The fourth feature matrix is ​​the product of the first feature matrix and the scaling scalar, so that the spatiotemporal relationship matrix, after controlling the intensity with the scaling scalar, is directly superimposed on the scaled dot product attention matrix to participate in the calculation of the attention weight matrix. Without performing high-dimensional linear transformation or dimensionality reduction through a multilayer perceptron on the spatiotemporal relationship matrix, the attention distribution between historical access interest points can be constrained during the self-attention calculation process. This makes the attention weight matrix contain both semantic relevance information based on the query matrix and key matrix and road network detour cost constraint information based on the fourth feature matrix, thereby reducing the computational complexity of the feature fusion process.

[0015] Furthermore, road network detour costs include: road network detour time costs and road network detour spatial distance costs.

[0016] Based on the aforementioned technical means, the road network detour cost includes two dimensions: road network detour time cost and road network detour spatial distance cost. This ensures that the road network detour cost between any two nodes in the spatiotemporal relationship matrix is ​​simultaneously constrained by two physical constraints: travel time and travel distance. This avoids recommendation bias caused by relying solely on a single cost dimension for route accessibility judgment. As a result, the target interest points determined based on the spatiotemporal relationship matrix have road network accessibility in both the actual travel time and actual distance dimensions, thereby improving the actual feasibility of the recommendation results under the current road network conditions.

[0017] Further, a basic feature matrix is ​​determined; wherein the basic feature matrix is ​​used to represent the access time characteristics and cabin status characteristics of each historical access point of interest in the historical access point of interest set; based on the historical access point of interest set, a first supplementary feature matrix is ​​determined; wherein the first supplementary feature matrix is ​​used to represent the historical access point of interest pairs with adjacent access times in the historical access point of interest set; based on the historical access point of interest set, a second supplementary feature matrix is ​​determined; wherein the second supplementary feature matrix is ​​used to represent the difference between the cabin status characteristics of each historical access point of interest in the historical access point of interest set and the current cabin status characteristics; based on the basic feature matrix, the first supplementary feature matrix, and the second supplementary feature matrix, a first feature matrix is ​​determined.

[0018] According to the above technical means, a basic feature matrix is ​​determined to represent the access time characteristics and cabin status characteristics of each historical access point of interest in the historical access point of interest set. A first supplementary feature matrix is ​​determined to represent the historical access point of interest pairs with adjacent access times in the historical access point of interest set. A second supplementary feature matrix is ​​determined to represent the difference between the cabin status characteristics of each historical access point of interest in the historical access point of interest set and the current cabin status characteristics. Based on the basic feature matrix, the first supplementary feature matrix, and the second supplementary feature matrix, a first feature matrix is ​​determined such that each historical access point of interest feature in the first feature matrix simultaneously includes the access time characteristics and cabin status characteristics of the historical access point of interest, the temporal correlation information between the historical access point of interest and adjacent access time points of interest, and the state difference information between the cabin status when the historical access point of interest occurred and the current cabin status.

[0019] Furthermore, based on the first supplementary feature matrix and the second supplementary feature matrix, the relative dynamic location index of each historical access point of interest is determined; wherein, the index number of the relative dynamic location index is used to indicate the access order of each historical access point of interest when revisiting it in the set of historical access points of interest; based on the relative dynamic location index and the basic feature matrix, the first feature matrix is ​​determined.

[0020] Based on the aforementioned technical means, the relative dynamic position index of each historical point of interest is determined by the first and second supplementary feature matrices. The index number of the relative dynamic position index is used to represent the access order of each historical point of interest revisited in the historical point of interest set. The first feature matrix is ​​determined based on the relative dynamic position index and the basic feature matrix, so that the position information of each historical point of interest in the sequence is no longer a fixed absolute sequence number, but a relative position code that is dynamically adjusted according to the adjacency relationship of access time and the difference in cabin status. This makes the position code of each historical point of interest in the first feature matrix change dynamically with the change of access time interval and cabin status difference, avoiding the loss of time reference value of fixed position code when the access time interval is uneven or the cabin status difference is large.

[0021] Furthermore, the access time characteristics include the access time and whether the access time is on a weekday; the cabin status characteristics include whether there are children in the cabin and the number of passengers in the cabin.

[0022] Based on the aforementioned technical means, by using access time characteristics (including access time and whether the access time is on a weekday) and cabin status characteristics (including whether there are children in the cabin and the number of passengers in the cabin), each historical access point of interest in the historical access point of interest set carries access time characteristics that include access time information and weekday / non-weekday distinction information, and cabin status characteristics that include whether there are children and the number of passengers. Thus, the historical access behavior data in the historical access point of interest set has the ability to distinguish between weekday and non-weekday patterns and the ability to perceive the cabin passenger status.

[0023] Furthermore, the candidate interest point set and travel intention vector are output by the large language model after the travel speech text is input; the user's recommendation feedback for the target interest point is determined; based on the recommendation feedback, the speech text recognition capability of the large language model is optimized.

[0024] Based on the aforementioned technical means, the large language model outputs travel speech text after inputting it into the candidate interest point set and travel intention vector. It then determines the user's recommendation feedback for the target interest point and optimizes the speech text recognition capability of the large language model based on the recommendation feedback. This allows the speech text recognition capability of the large language model to be continuously optimized according to the user's actual selection behavior of the recommendation results. That is, positive feedback from the user to the target interest point strengthens the large language model's recognition preference for similar travel speech texts, while negative feedback from the user to the target interest point inhibits the large language model's recognition tendency for similar travel speech texts. After multiple recommendation feedback loops, the large language model gradually converges to personalized recognition parameters based on the current user's travel speech expression habits.

[0025] Secondly, this application provides a travel recommendation device, including a determining module, a generating module, and a recommending module. The determining module is used to determine a set of candidate points of interest based on the user's travel voice text. The generating module is used to generate a set of historical points of interest based on the user's historical access points of interest. The historical points of interest in the historical access point set include access time features and cabin status features. The determining module is also used to determine a spatiotemporal relationship matrix based on the candidate point of interest set and the historical point of interest set. The spatiotemporal relationship matrix is ​​used to characterize the road network detour cost between a first target node and a second target node. The first target node is a node in the candidate point of interest set. The second target node is a node in the historical point of interest set. The recommending module is used to recommend travel points of interest to the user based on the spatiotemporal relationship matrix, the historical point of interest set, and the candidate point of interest set.

[0026] Furthermore, the determination module is specifically used to determine the first feature matrix corresponding to the set of historical access points of interest; based on the spatiotemporal relationship matrix, attention learning is performed on the first feature matrix to obtain the second feature matrix. The attention of the target historical access points of interest in the first feature matrix is ​​determined based on the road network detour cost between the target historical access points of interest and the target candidate points of interest; the target candidate points of interest are candidate points of interest in the spatiotemporal relationship matrix that have a road network detour relationship with the target historical access points of interest; based on the set of candidate points of interest and the second feature matrix, the target points of interest are determined; and the target points of interest are recommended to the user.

[0027] Furthermore, the determination module is specifically used to determine the travel intention vector based on travel voice text; based on the travel intention vector, to perform attention learning on the second feature matrix to determine the third feature matrix; and based on the candidate interest point set and the third feature matrix, to determine the target interest point.

[0028] Furthermore, the determination module is specifically used to determine the travel score of each candidate interest point in the candidate interest point set based on the third feature matrix; and to determine the first preset number of candidate interest points ranked by travel score as the target interest points.

[0029] Furthermore, the determination module is specifically used to determine the query matrix, key matrix, and value matrix corresponding to the first feature matrix; determine the scaled dot product attention matrix of the query matrix and key matrix; determine the attention weight matrix based on the fusion matrix of the scaled dot product attention matrix and the fourth feature matrix; wherein the fourth feature matrix is ​​the product of the first feature matrix and the scaling scalar; and determine the second feature matrix based on the attention weight matrix and the value matrix.

[0030] Furthermore, road network detour costs include: road network detour time costs and road network detour spatial distance costs.

[0031] Further, the determination module is specifically used to determine the basic feature matrix; wherein the basic feature matrix is ​​used to represent the access time characteristics and cabin status characteristics of each historical access point of interest in the historical access point of interest set; based on the historical access point of interest set, a first supplementary feature matrix is ​​determined; wherein the first supplementary feature matrix is ​​used to represent the historical access point of interest pairs with adjacent access times in the historical access point of interest set; based on the historical access point of interest set, a second supplementary feature matrix is ​​determined; wherein the second supplementary feature matrix is ​​used to represent the difference between the cabin status characteristics of each historical access point of interest in the historical access point of interest set and the current cabin status characteristics; based on the basic feature matrix, the first supplementary feature matrix, and the second supplementary feature matrix, a first feature matrix is ​​determined.

[0032] Furthermore, the determining module is specifically used to determine the relative dynamic location index of each historical access point of interest based on the first supplementary feature matrix and the second supplementary feature matrix; wherein, the index number of the relative dynamic location index is used to indicate the access order of each historical access point of interest when revisiting it in the set of historical access points of interest; and to determine the first feature matrix based on the relative dynamic location index and the basic feature matrix.

[0033] Furthermore, the access time characteristics include the access time and whether the access time is on a weekday; the cabin status characteristics include whether there are children in the cabin and the number of passengers in the cabin.

[0034] Furthermore, the candidate interest point set and travel intention vector are output by the large language model after the travel speech text is input; the determination module is specifically used to determine the user's recommendation feedback for the target interest point; based on the recommendation feedback, the speech text recognition capability of the large language model is optimized.

[0035] Thirdly, this application provides an electronic device comprising: a processor and a memory; the memory storing processor-executable instructions. When the processor is configured to execute the instructions, the electronic device implements the method described in the first aspect.

[0036] Fourthly, this application provides a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the method described in the first aspect.

[0037] Fifthly, this application provides a computer program product including computer program instructions that, when executed by a processor, implement the method described in the first aspect.

[0038] Sixthly, this application provides a vehicle that makes travel recommendations using the travel recommendation method provided in the first aspect above.

[0039] It should be noted that the technical effects of any of the implementation methods in aspects two through six can be found in the technical effects of the corresponding implementation methods in aspect one, and will not be repeated here.

[0040] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the architecture of a travel recommendation system provided by the present invention; Figure 2 A flowchart illustrating a travel recommendation method provided by the present invention; Figure 3 This is a flowchart illustrating another travel recommendation method provided by the present invention. Figure 4 This is a schematic diagram of the composition of a travel recommendation device provided by the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device provided by the present invention. Detailed Implementation

[0042] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0043] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0044] The following is a detailed description of a travel recommendation method, device, and vehicle provided in this application, with reference to the accompanying drawings.

[0045] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0046] The terms "first" and "second," etc., used in the specification and drawings of this application are used to distinguish different objects or to distinguish different treatments of the same object, rather than to describe a specific order of objects.

[0047] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0048] It should be noted that in the embodiments of this application, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplarily" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0049] To facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with essentially the same function and effect. Those skilled in the art can understand that the terms "first" and "second" are not intended to limit the quantity or execution order.

[0050] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0051] The embodiments provided in this application will now be described in detail with reference to the accompanying drawings.

[0052] The travel recommendation method provided in this application can be applied to, for example... Figure 1 The travel recommendation system shown. For example... Figure 1 As shown, the travel recommendation system 10 of this application includes: a backend device 11 and a frontend device 12.

[0053] A communication connection is established between the backend device 11 and the frontend device 12. For example, the connection method can be wireless, such as Bluetooth or Wi-Fi; or it can be wired, such as fiber optic, etc., without limitation. For example, the backend device 11 and the frontend device 12 can be connected to the Internet via a router, thereby enabling the communication connection between them.

[0054] In some embodiments, the backend device 11 may receive the user's travel voice text sent by the frontend device 12, so as to determine a set of candidate points of interest based on the received travel voice text.

[0055] In some embodiments, the backend device 11 can also determine a spatiotemporal relationship matrix based on the candidate point of interest set and the historical access point of interest set, and further recommend travel points of interest to users based on the spatiotemporal relationship matrix, the historical access point of interest set and the candidate point of interest set.

[0056] In some embodiments, the backend device 11 may also send the recommended points of interest to the frontend device 12 so that the frontend device 12 can display the points of interest to the user.

[0057] In some embodiments, the backend device 11 can be a server, such as a single server or a server cluster consisting of multiple servers. In some embodiments, the server cluster can also be a distributed cluster.

[0058] In some embodiments, the backend device 11 can be a terminal device, such as a mobile phone, tablet computer, desktop computer, laptop computer, handheld computer, notebook computer, in-vehicle processor, ultra-mobile personal computer (UMPC), netbook, cellular phone, personal digital assistant (PDA), augmented reality (AR) / virtual reality (VR) device, etc. This application embodiment does not impose any special limitations on the specific form of the terminal device.

[0059] In some embodiments, the front-end device 12 is used to display recommended points of interest. For example, the prompting device 12 may be a voice prompting device, in which case the prompting device 12 presents the recommended points of interest to the user by reading them aloud. Alternatively, the prompting device 12 may also be a display device, in which case the prompting device 12 presents the recommended points of interest to the user by displaying them on a display screen.

[0060] For example, the front-end device 12 can be an in-vehicle central control display screen, which displays recommended point-of-interest (POI) cards through a graphical user interface. The POI cards include at least the POI name, distance, estimated detour time, and star rating information. The front-end device 12 can also be a head-up display (HUD), which projects the names and directional information of the recommended POIs onto the vehicle's windshield.

[0061] In some embodiments, the front-end device 12 may also be a mobile terminal device, such as a mobile phone, tablet, or smartwatch. The mobile terminal device communicates and connects with the back-end device via Bluetooth, Wi-Fi, or vehicle networking to display recommended points of interest information after the user leaves the vehicle.

[0062] In some embodiments, the electronic device 11 and the storage device 12 can be as follows: Figure 1 As shown, there are two separate devices, or electronic device 11 and storage device 12 can be integrated into the same device.

[0063] It should be noted that the system architecture described in the embodiments of this application is for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and does not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of system architecture, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0064] The travel recommendation method provided in this application can be applied to travel recommendation systems, such as... Figure 2 As shown, the travel recommendation method includes the following steps S201~S204: S201. Determine a set of candidate points of interest based on the user's travel voice text.

[0065] In some embodiments, a user's travel voice text refers to a natural language statement uttered by the user via voice input, containing travel intentions or point-of-interest queries. Travel voice text is the text form of the received raw speech signal converted through speech recognition, carrying descriptive information about the user's destination type, location range, functional requirements, or other preferences. Travel voice text can be a complete statement spoken by the user, such as "I want to go to a nearby gas station," or a complex query statement containing multiple conditions or modifiers, such as "Find me a Sichuan restaurant nearby with parking that is open until 10 PM."

[0066] In some embodiments, the candidate point of interest (POI) set refers to a set of one or more POIs initially selected by the system based on the user's travel voice text. Each POI in the candidate POI set has a potential semantic relevance to the user's travel intent, serving as a candidate for subsequent fine-tuning and recommendation processes. The candidate POI set can be generated by semantically understanding the travel voice text using a large language model, or it can be recalled by keyword matching of POI names and category tags. The candidate POI set includes one or more basic attribute information such as POI name, POI category, POI location coordinates, POI address, POI contact information, and POI operating status.

[0067] In some embodiments, in response to a user triggering a voice acquisition operation, the system acquires the user's original voice signal through an in-vehicle microphone array, performs analog-to-digital conversion on the original voice signal, extracts features from the converted digital voice signal to obtain a voice feature sequence, inputs the voice feature sequence into a pre-trained acoustic model for phoneme recognition, and decodes the text based on the recognized phoneme sequence through a language model to output the user's travel voice text.

[0068] In some embodiments, the acoustic model can be pre-trained as follows: a large amount of speech sample data and the corresponding labeled text for each speech sample data are acquired; acoustic features are extracted from the speech sample data to obtain a speech feature sequence; the speech feature sequence is used as the input of the acoustic model; the labeled text is used as the supervision label; the network parameters of the acoustic model are iteratively adjusted through the backpropagation algorithm until the recognition accuracy of the acoustic model on the validation set reaches a preset threshold, thus obtaining the trained acoustic model.

[0069] In some embodiments, after acquiring the user's travel voice text, the travel voice text is input into a large language model for semantic parsing, and the large language model outputs a structured intent. After acquiring the structured intent, the target description field and category field are read from the structured intent, and keywords from the target description field and category field are extracted as search terms. At the same time, the spatial constraint field in the structured intent is read, and the geofence radius and spatial reference point are extracted from the spatial constraint field. Using the vehicle's current location as the spatial reference point and a preset radius as the geofence radius, all points of interest whose location coordinates fall within the geofence range are retrieved from the vehicle POI database to obtain a preliminary spatial candidate set. Using the search terms as query conditions, inverted index matching or vector similarity retrieval is performed on the preliminary spatial candidate set. The matching score between each point of interest in the preliminary spatial candidate set and the search term is calculated. The points of interest in the preliminary spatial candidate set are sorted in descending order of matching score, and the top M points of interest in the sorted results are selected as a coarse-screened candidate POI set, where M is a preset positive integer.

[0070] In some embodiments, the training of a large language model includes a pre-training phase and a fine-tuning phase. In the pre-training phase, unlabeled text corpora are acquired as pre-training sample data. This pre-training sample data is then vectorized. The vectorized text sequence is input into an initial language model. The initial language model predicts lexical units in the text sequence based on an autoregressive or autoencoder pre-training task. The predicted results are compared with the actual lexical units in the text sequence. The pre-training loss value is calculated based on the comparison results. The network parameters of the initial language model are iteratively adjusted using a backpropagation algorithm until the pre-training loss value converges to a preset range, thus obtaining the pre-trained language model. In the fine-tuning phase, fine-tuning sample data labeled with task tags is acquired. This fine-tuning sample data is input into the pre-trained language model. The pre-trained language model outputs prediction results for the fine-tuning task. The prediction results are compared with the task tags. The fine-tuning loss value is calculated based on the comparison results. The network parameters of the pre-trained language model are iteratively adjusted using a backpropagation algorithm until the fine-tuning loss value converges to a preset range, thus obtaining the trained large language model.

[0071] In some embodiments, to address the illusion problem that large language models are prone to and to ensure the standardization of the output format, a cue word engineering approach combined with thought chains is employed. The cue words injected into the large language model predefine an atomic capability tag library (such as category, taste, facility, distance constraint, etc.) for in-vehicle POI recommendations.

[0072] For example, the large language model infers from the instruction text and outputs a structured semantic intent feature dictionary, which can be: Intent = {Category: "catering", Tags: ["family", "light", "play area"], Spatial_Constraint: "on the way / ahead"}.

[0073] S202. Generate a set of historical access points of interest based on the user's historical access points of interest.

[0074] The historical access points of interest set includes access time characteristics and cabin status characteristics.

[0075] In some embodiments, historical points of interest (POIs) refer to records of points of interest that a user has initiated, completed, or actively saved through the vehicle's infotainment system during past trips. Historical POIs reflect a user's point-of-interest selection behavior and preferences in real-world driving scenarios. They are important nodes in a user's historical travel trajectory and generate genuine interaction records with the user's actual driving behavior, including instances where the user set the POI as a navigation destination, a transit point, or searched for or viewed details of the POI.

[0076] For example, each historical point of interest can be represented as a tuple. .in, Feature embeddings representing historical visit points of interest (e.g., the type of point of interest). Representative access time characteristics, This represents the cabin status characteristics at the time the visit occurred.

[0077] In some embodiments, a historical access point of interest set refers to a data set consisting of one or more historical access points of interest. Each historical access point of interest in the historical access point of interest set corresponds to one historical access behavior record, and all historical access behavior records are arranged in chronological order according to the access timestamps, forming the user's historical travel sequence. The historical access point of interest set covers the user's access trajectory of points of interest over a past period of time. This period of time can be a preset time window length, such as the past three months or the past year, or a preset access frequency window, such as the most recent one hundred access records.

[0078] In some embodiments, the access time feature refers to feature data that characterizes the time information of a user's access to a point of interest. The access time feature may include the access time and whether the access time is on a weekday.

[0079] In some embodiments, cabin state characteristics refer to feature data characterizing the environmental and occupant states within the vehicle cabin at the time of a point-of-interest visit. Cabin state characteristics may include the presence of children in the cabin and the number of occupants in the cabin.

[0080] For example, cabin state characteristics acquired through cameras from the driver monitoring system, passenger monitoring system, and seat gravity sensors deployed within the cabin can be characterized as follows: ,in, This indicates whether a child safety seat is occupied in the cabin. This indicates the total number of occupants in the cabin.

[0081] S203. Determine the spatiotemporal relationship matrix based on the candidate interest point set and the historical access interest point set.

[0082] The spatiotemporal relationship matrix is ​​used to characterize the road network detour cost between the first target node and the second target node; the first target node is a node in the candidate point of interest set; and the second target node is a node in the historical visited point of interest set.

[0083] In some embodiments, the spatiotemporal relationship matrix refers to a matrix whose row and column nodes are each point of interest in the candidate point of interest set and each point of interest in the historical visited point of interest set. The matrix element values ​​represent the detour cost added to the road network after adding the corresponding two nodes. The number of rows in the spatiotemporal relationship matrix is ​​equal to the sum of the number of points of interest in the candidate point of interest set and the number of points of interest in the historical visited point of interest set, and the number of columns is also equal to the sum of the number of points of interest in the candidate point of interest set and the number of points of interest in the historical visited point of interest set.

[0084] In some embodiments, road network detour cost refers to the additional time and distance cost incurred by a vehicle when a node in the matrix is ​​added to a route that the vehicle is currently executing or has already planned. Specifically, detour cost represents the incremental cost of the actual travel path relative to the original planned path after adding the point of interest corresponding to the node to a given travel route.

[0085] In some embodiments, road network detour costs include: road network detour time costs and road network detour spatial distance costs.

[0086] In some embodiments, road network detour time cost refers to the additional travel time required for a vehicle when the system adds nodes from the matrix to a route that the vehicle is currently executing or has already planned.

[0087] In some embodiments, the road network detour space distance cost refers to the additional travel distance required by a vehicle when the system adds a node in the matrix to the vehicle's current or planned travel route. The travel distance is the actual distance traveled to the node, not the straight-line distance.

[0088] For example, after adding each POI node from the candidate point of interest set and each POI from the historical access point of interest set to the road network, the route is replanned, and a path planning API call is initiated to the underlying in-vehicle high-precision map or standard navigation engine. The navigation engine calculates the road network detour time cost and road network detour spatial distance cost based on real-time traffic conditions (congestion) and road topology (one-way streets, overpass restrictions, no U-turn constraints).

[0089] For example, the parameters passed to the navigation engine may also include: the vehicle's current position and the current heading angle. When the vehicle has a predetermined navigation route, the onboard navigation engine takes the vehicle's current position as the starting point and the destination of the original navigation route as the ending point, obtains the total travel time and total travel distance after replanning, subtracts the total travel time of the original navigation route from the replanned total travel time to obtain the corresponding road network detour time cost, and subtracts the total travel distance of the original navigation route from the replanned total travel distance to obtain the corresponding road network detour spatial distance cost. When the vehicle does not have a predetermined navigation route, the onboard navigation engine takes the vehicle's current position as the starting point, obtains the estimated travel time and estimated travel distance from the vehicle's current position to the node, determines the estimated travel time as the corresponding road network detour time cost, and determines the estimated travel distance as the corresponding road network detour spatial distance cost.

[0090] In some embodiments, to prevent abnormal extreme values ​​in the spatiotemporal relationship matrix from interfering with the gradient calculation of the neural network, maximum time thresholds and maximum distance thresholds can be set for the time interval and spatial detour cost, respectively.

[0091] In some embodiments, the detour cost exceeding the threshold can be truncated. That is, when the detour cost is greater than the maximum time threshold or the maximum distance threshold, the maximum time threshold is used to replace the detour time cost, and the maximum distance threshold is used to replace the detour spatial distance cost in the detour cost.

[0092] In some embodiments, a spatiotemporal relationship score can be calculated based on a maximum time threshold, a maximum distance threshold, and a road network detour cost, and the spatiotemporal relationship score can be used as an element in a matrix.

[0093] For example, the method for calculating the spatiotemporal relationship score can satisfy the following relationship:

[0094] in, Indicates the score for spatiotemporal relationship. Indicates the maximum time threshold. Represents a node and The corresponding road network detour time cost, Indicates the maximum distance threshold. Indicates the cost of road network detour space. and These represent time-based weighting and space-based weighting, respectively.

[0095] S204. Based on the spatiotemporal relationship matrix, the set of historical access points of interest, and the set of candidate points of interest, recommend travel points of interest to users.

[0096] In some embodiments, recommending travel points of interest to users can be implemented through the following steps a1~a4: a1. Determine the first feature matrix corresponding to the set of historical visit points of interest.

[0097] In some embodiments, the vectors in the first feature matrix all incorporate the original semantic information of the set of historical access points of interest, the time span features between historical access points of interest, and the cabin scene features when the access to the historical access points of interest occurred.

[0098] The following section introduces the method for determining the first characteristic matrix.

[0099] The first step is to determine the basic feature matrix; whereby the basic feature matrix is ​​used to represent the access time characteristics and cabin status characteristics of each historical access point of interest in the historical access point of interest set.

[0100] In some embodiments, the access time characteristics and cabin status characteristics of each historical access point of interest can be determined in the manner described in step S202 to form a basic feature matrix.

[0101] The second step is to determine the first supplementary feature matrix based on the set of historical access points of interest; wherein the first supplementary feature matrix is ​​used to represent the pairs of historical access points of interest with adjacent access times in the set of historical access points of interest.

[0102] For example, for the Kth visit and the K+1th visit in the sequence, the absolute time interval between the two can be used as an element in the first supplementary feature matrix.

[0103] The third step is to determine the second supplementary feature matrix based on the set of historical access points of interest; wherein the second supplementary feature matrix is ​​used to represent the difference between the cabin state features of each historical access point of interest in the set of historical access points of interest and the current cabin state features.

[0104] For example, the second supplementary feature matrix can be determined using a variant of cosine distance or weighted Euclidean distance, and the second supplementary feature matrix can satisfy the following relationship:

[0105] in, This represents the element corresponding to the k-th historical visit point of interest in the second supplementary feature matrix. This represents the cockpit state characteristics when the k-th historical point of interest is visited. This represents the current cabin state characteristics. If only the driver was in the car during a previous visit (such as going to a bar), but the current cabin state detects a child in the car, the state vectors of the two will differ significantly. Approaching 1 (or setting an upper limit); conversely, if the cockpit states of both are highly consistent, Approaching 0.

[0106] The fourth step is to determine the first feature matrix based on the basic feature matrix, the first supplementary feature matrix, and the second supplementary feature matrix.

[0107] In some embodiments, the relative dynamic location index of each historical access point of interest is determined based on the first supplementary feature matrix and the second supplementary feature matrix; wherein, the index number of the relative dynamic location index is used to indicate the access order of each historical access point of interest when revisiting it in the set of historical access points of interest.

[0108] In some embodiments, the relative dynamic location index is used to represent the index number of the revisited order of the historically visited point of interest in the set of historically visited points of interest. The relative dynamic location index differs from the traditional fixed location index, which assigns an increasing sequence number solely based on the order of access timestamps. In contrast, the relative dynamic location index integrates information about the time interval between adjacent visits, cabin status differences, and cabin status similarity. This results in visits with similar times and cabin status having similar index numbers in the index space, while visits with large time spans or significant cabin status differences have index numbers that are far apart in the index space.

[0109] In some embodiments, the index number is a non-negative integer value. When the index numbers are arranged in ascending order, the closer the index numbers of two historical access points of interest, the higher the coherence and similarity of these two access behaviors in the time dimension and the cabin scene dimension. The farther apart the index numbers of two historical access points of interest, the more significant the jump or change in the time dimension or the cabin scene dimension of these two access behaviors.

[0110] For example, the relative dynamic position index of the (k+1)th access in the sequence can be calculated recursively based on the starting position:

[0111] in, This represents the relative dynamic position index of the (k+1)th historical point of interest in the sequence. This represents the relative dynamic position index of the k-th historical point of interest in the sequence. The absolute time interval between the k-th historical point of interest and the (k+1)-th historical point of interest can be obtained from the first supplementary feature matrix. This represents the average of all absolute time intervals in the set of historical access points of interest. This represents the element corresponding to the (k+1)th historical visit point of interest in the second supplementary feature matrix, which is the difference penalty for the cabin state feature. This indicates the assignment of weights to absolute time intervals. This indicates the penalty allocation weight, and the constant term 1 is used to ensure that even if the time interval is extremely short and the scenario is the same, the two visits still have the minimum basic position difference.

[0112] In some embodiments, after determining the relative dynamic position index, the first feature matrix can be determined based on the relative dynamic position index and the basic feature matrix.

[0113] In some embodiments, all the calculated dynamic position indices are mapped to position encoding vectors using a sine and cosine transform function.

[0114] For example, a function mapping can satisfy the following relationship:

[0115]

[0116] in, This represents the relative dynamic position index of the k-th historical point of interest in the sequence. This represents the dimension index of the position encoding vector. Represents the dimension of a vector. This represents the relative dynamic position index of the k-th historical access point of interest in the sequence. A positional encoding vector with 3 dimensions. This represents the relative dynamic position index of the k-th historical access point of interest in the sequence. A positional encoding vector with multiple dimensions.

[0117] In some embodiments, the position encoding vector is injected into the base feature matrix by adding elements one by one to obtain the first feature matrix.

[0118] In some embodiments, the first feature matrix can, before entering the deep network, bring together relevant historical memories (such as playgrounds and family restaurants that have been visited in the past) that will be awakened by the current cabin state, such as "children riding in a car", in the shallow feature space, giving them higher prior weights.

[0119] a2. Based on the spatiotemporal relationship matrix, attention learning is performed on the first feature matrix to obtain the second feature matrix. The attention of the target historical visit interest points in the first feature matrix is ​​determined based on the road network detour cost between the target historical visit interest points and the target candidate interest points. The target candidate interest points are the candidate interest points in the spatiotemporal relationship matrix that have a road network detour relationship with the target historical visit interest points.

[0120] In some embodiments, attention learning refers to the process of calculating the attention weights between historical points of interest in a first feature matrix based on the road network detour costs in the spatiotemporal relationship matrix, and then weighting and aggregating the first feature matrix according to the attention weights.

[0121] In some embodiments, to learn the detour features in both time and space, it is typically necessary to map the spatiotemporal relationship matrix into a high-dimensional tensor and concatenate it with the feature vectors, followed by dimensionality reduction through a multi-layer fully connected network. This results in a quadratic increase in the number of model parameters, placing catastrophic pressure on the limited computing power of the vehicle. This application proposes a lightweight inductive bias injection method that is highly computationally efficient to determine the second feature matrix.

[0122] In some embodiments, the method for determining the second feature matrix is ​​described below: The first step is to determine the query matrix, key matrix, and value matrix corresponding to the first feature matrix.

[0123] In some embodiments, the first feature matrix is ​​mapped to a query matrix, a key matrix, and a value matrix through three different learnable linear weight matrices.

[0124] The second step is to determine the scaling dot product attention matrix for the query matrix and the key matrix.

[0125] For example, it can be done by To determine the scaled dot product attention matrix, where, Represents the query matrix. Represents the key matrix, Indicates the dimension of the vector.

[0126] In some embodiments, scaling the dot product attention matrix can leverage neural networks to uncover the intrinsic associations of sequences in the latent semantic space (e.g., if a user frequently visits a Sichuan restaurant, the semantic association is high even if the restaurant is far away).

[0127] The third step is to determine the attention weight matrix based on the fusion matrix of the scaled dot product attention matrix and the fourth feature matrix; where the fourth feature matrix is ​​the product of the first feature matrix and the scaling scalar.

[0128] For example, the fourth feature matrix is ,in, Indicates scaling of scalar. This represents the first characteristic matrix.

[0129] For example, the attention weight matrix can be .

[0130] The fourth step is to determine the second feature matrix based on the attention weight matrix and the value matrix.

[0131] For example, the second feature matrix can be determined in a way that satisfies the following relationship:

[0132] in, This represents the second characteristic matrix. Represents the attention weight matrix. Represents a value matrix, This represents the normalization function.

[0133] In some embodiments, the second feature matrix obtained by multi-layer attention computation (typically stacked 2-4 layers with residual connections introduced between layers) can extract feature sequences from the set of historical access points of interest that contain drivers’ long-term habits and conform to the characteristics of the current road network.

[0134] a3. Based on the candidate interest point set and the second feature matrix, determine the target interest point.

[0135] In some embodiments, a target point of interest refers to one or more points of interest that are ultimately selected from the set of candidate points of interest and recommended to the user.

[0136] The following describes the methods for determining target points of interest: The first step is to determine the travel intent vector based on the travel voice text.

[0137] In some embodiments, a structured semantic intent feature dictionary extracted from a large language model is mapped into a high-dimensional travel intent vector through a lightweight text embedding model (such as RoBERTa).

[0138] The second step involves performing attention learning on the second feature matrix based on the travel intention vector to determine the third feature matrix.

[0139] In some embodiments, the intent vector can be used as the query subject, the second feature matrix can be used as the key subject and value subject to be retrieved, and the third feature matrix can be determined.

[0140] For example, the third feature matrix can be determined to satisfy the following relationship:

[0141] in, Represents the third characteristic matrix. Indicates the subject of the query. This indicates the transpose of the key body. Indicates the value body, Represents the dimension of a vector. This represents the normalization function.

[0142] For example, when the big parser detects the intent of "suitable for children", it automatically finds those access records with "parent-child" attributes in the second feature matrix and assigns them a very high attention weight. Finally, its features are extracted and condensed into the third feature matrix, which makes up for the lack of deep user history memory on the big model in the car.

[0143] The third step is to determine the target interest point based on the candidate interest point set and the third feature matrix.

[0144] In some embodiments, the travel score for each candidate point of interest in the candidate point of interest set is determined based on a third feature matrix.

[0145] In some embodiments, the travel score can be output by performing an inner product calculation on the candidate interest point set and the third feature matrix one by one.

[0146] For example, the travel score for each candidate point of interest in the candidate point of interest set can satisfy the following relationship:

[0147] in, Represents the interest points in the candidate interest point set. Travel score, Represents the third characteristic matrix. Represents the interest points in the candidate interest point set. The embeddings of interest points in the candidate interest point set can be mapped to embeddings using a lightweight text embedding model. express and Perform matrix multiplication operations between them.

[0148] In some embodiments, a preset number of candidate points of interest ranked first by travel score are determined as target points of interest.

[0149] In some embodiments, the travel score comprehensively measures the overall performance of candidate points of interest across four dimensions: “meeting the user’s verbal intent,” “aligning with the user’s historical taste preferences,” “fitting the current physical cabin environment configuration,” and “following the route without detours.”

[0150] a4. Recommend target points of interest to users.

[0151] In some embodiments, the trip scores can be sorted in descending order, low-scoring items can be removed, and the K (e.g., K=3) candidate locations with the highest confidence levels can be selected and assembled into a recommendation card data package containing the POI name, distance, estimated detour time, and star rating, which is then displayed to the user. For example, the recommendation card data package can be transmitted back to the cockpit operating system via the vehicle bus and rendered on the central touchscreen, passenger entertainment screen, or head-up display interface.

[0152] In some embodiments, the travel recommendation method provided in this application further includes: determining the user's recommendation feedback on the target point of interest; and optimizing the speech and text recognition capabilities of a large language model based on the recommendation feedback.

[0153] For example, if a user clicks on a recommended point of interest (POI) and successfully initiates navigation, that POI is marked as a positive sample and counted as a hit. If the user ignores the recommended POI, manually cancels, or initiates a new round of precise search after a voice interruption, all currently recommended POIs are marked as hard negative samples. Based on positive and hard negative samples, low-rank adaptive fine-tuning techniques or automatic optimization of prompt words can be used to continuously improve the recall rate of the large language model in the initial semantic screening stage for recognizing user intent.

[0154] For example, Figure 3 This is a flowchart illustrating another travel recommendation method provided by the present invention. Figure 3As shown, after the process begins, the following steps are executed: Step S1: Obtain the user's travel voice text. Step S2: Intent classification and recognition. Step S3: Determine if it is an interest point recommendation task. If not, execute Step S4: Terminate the current process; if yes, execute Step S5: Collect intent, vehicle condition, cabin status, and time information. Step S6: Preprocess the intent using a large language model. Step S7: Extract the intent vector and determine the candidate interest point set. Step S8: Determine if historical access interest point information is contained. If not, execute Step S9: Score the candidate interest points semantically using a large oracle model, then execute Step S10: Generate recommendation cards; if yes, execute Step S11: Obtain the historical access interest point set. Step S12: Determine the relative dynamic location index of each historical access interest point. Step S13: Construct a spatiotemporal relationship matrix. Step S14: Determine the third feature matrix. Step S15: Based on the third feature matrix, determine the travel score for each candidate interest point. Step S16: Generate recommendation cards based on travel scores. Step S17: Construct the card format to be passed to the vehicle's infotainment system. Step S18: Transmit card data. Step S19: Determine if validity and domain validation pass. If not, proceed to step S20 and return an error message; if yes, proceed to step S21 and edit the card data. Step S22: Send the results to the vehicle's infotainment system. Step S23: Implement a pop-up window for user-initiated navigation. Step S24: Optimize the large language model based on recommendation feedback. Finally, the process ends.

[0155] The foregoing mainly describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the above functions, the travel recommendation device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0156] like Figure 4As shown, the travel recommendation device 1000 includes a determining module 1001, a generating module 1002, and a recommending module 1003. The determining module 1001 is used to determine a set of candidate points of interest based on the user's travel voice text. The generating module 1002 is used to generate a set of historical points of interest based on the user's historical access points of interest. The historical points of interest in the historical access point of interest set include access time characteristics and cabin status characteristics. The determining module 1001 is also used to determine a spatiotemporal relationship matrix based on the candidate point of interest set and the historical point of interest set. The spatiotemporal relationship matrix is ​​used to characterize the road network detour cost between a first target node and a second target node. The first target node is a node in the candidate point of interest set. The second target node is a node in the historical point of interest set. The recommending module 1003 is used to recommend travel points of interest to the user based on the spatiotemporal relationship matrix, the historical point of interest set, and the candidate point of interest set.

[0157] Furthermore, the determination module 1001 is specifically used to determine the first feature matrix corresponding to the set of historical access points of interest; based on the spatiotemporal relationship matrix, attention learning is performed on the first feature matrix to obtain the second feature matrix. The attention of the target historical access points of interest in the first feature matrix is ​​determined based on the road network detour cost between the target historical access points of interest and the target candidate points of interest; the target candidate points of interest are candidate points of interest in the spatiotemporal relationship matrix that have a road network detour relationship with the target historical access points of interest; based on the set of candidate points of interest and the second feature matrix, the target points of interest are determined; and the target points of interest are recommended to the user.

[0158] Furthermore, the determination module 1001 is specifically used to determine the travel intention vector based on the travel voice text; to perform attention learning on the second feature matrix based on the travel intention vector to determine the third feature matrix; and to determine the target interest point based on the candidate interest point set and the third feature matrix.

[0159] Furthermore, the determination module 1001 is specifically used to determine the travel score of each candidate interest point in the candidate interest point set based on the third feature matrix; and to determine the preset number of candidate interest points with the highest travel scores as target interest points.

[0160] Furthermore, the determining module 1001 is specifically used to determine the query matrix, key matrix, and value matrix corresponding to the first feature matrix; determine the scaled dot product attention matrix of the query matrix and key matrix; determine the attention weight matrix based on the fusion matrix of the scaled dot product attention matrix and the fourth feature matrix; wherein the fourth feature matrix is ​​the product of the first feature matrix and the scaling scalar; and determine the second feature matrix based on the attention weight matrix and the value matrix.

[0161] Furthermore, road network detour costs include: road network detour time costs and road network detour spatial distance costs.

[0162] Further, the determining module 1001 is specifically used to determine the basic feature matrix; wherein the basic feature matrix is ​​used to represent the access time characteristics and cabin status characteristics of each historical access point of interest in the historical access point of interest set; based on the historical access point of interest set, a first supplementary feature matrix is ​​determined; wherein the first supplementary feature matrix is ​​used to represent the historical access point of interest pairs with adjacent access times in the historical access point of interest set; based on the historical access point of interest set, a second supplementary feature matrix is ​​determined; wherein the second supplementary feature matrix is ​​used to represent the difference between the cabin status characteristics of each historical access point of interest in the historical access point of interest set and the current cabin status characteristics; based on the basic feature matrix, the first supplementary feature matrix, and the second supplementary feature matrix, a first feature matrix is ​​determined.

[0163] Furthermore, the determining module 1001 is specifically used to determine the relative dynamic location index of each historical access point of interest based on the first supplementary feature matrix and the second supplementary feature matrix; wherein, the index number of the relative dynamic location index is used to indicate the access order of each historical access point of interest when revisiting it in the set of historical access points of interest; and to determine the first feature matrix based on the relative dynamic location index and the basic feature matrix.

[0164] Furthermore, the access time characteristics include the access time and whether the access time is on a weekday; the cabin status characteristics include whether there are children in the cabin and the number of passengers in the cabin.

[0165] Furthermore, the candidate interest point set and travel intention vector are output by the large language model after the travel speech text is input; the determination module 1001 is specifically used to determine the user's recommendation feedback for the target interest point; based on the recommendation feedback, the speech text recognition capability of the large language model is optimized.

[0166] like Figure 5 As shown, the electronic device 1100 includes, but is not limited to, a processor 1101 and a memory 1102.

[0167] The aforementioned memory 1102 is used to store the executable instructions of the processor 1101. It is understood that the processor 1101 is configured to execute instructions to implement the travel recommendation method in the above embodiments.

[0168] It should be noted that those skilled in the art will understand that Figure 5 The electronic device structure shown does not constitute a limitation on electronic device 1100; electronic device 1100 may include, but is not limited to, other electronic devices. Figure 5This may indicate more or fewer components, or combinations of certain components, or different component arrangements.

[0169] Processor 1101 is the control center of electronic device 1100. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in memory 1102, and by calling data stored in memory 1102, it performs various functions and processes data of electronic device 1100, thereby providing overall monitoring of electronic device 1100. Processor 1101 may include one or more processing units. Optionally, processor 1101 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into processor 1101.

[0170] The memory 1102 can be used to store software programs and various data. The memory 1102 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required by at least one functional module (such as a determination unit, processing unit, etc.), etc. Furthermore, the memory 1102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0171] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 1102 including instructions, which can be executed by a processor 1101 of an electronic device 1100 to implement the travel recommendation method in the above embodiments.

[0172] In actual implementation, Figure 4 The functions of the determination module 1001, generation module 1002, and recommendation module 1003 can all be provided by... Figure 5 The processor 1101 calls the computer program stored in the memory 1102 to implement the process. The specific execution process can be found in the method section of the previous embodiment, and will not be repeated here.

[0173] Optionally, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0174] In an exemplary embodiment, this application also provides a computer program product including one or more instructions, which can be executed by the processor 1101 of the electronic device 1100 to complete the travel recommendation method in the above embodiments.

[0175] It should be noted that when one or more instructions in the computer-readable storage medium or computer program product are executed by the processor of an electronic device, they implement the various processes of the above method embodiments and achieve the same technical effect as the above method. To avoid repetition, they will not be described again here.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0177] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another apparatus, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0178] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units; that is, it can be located in one place or distributed in multiple different locations. Some or all of the classified units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0179] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0180] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, essentially, or the part that contributes to the prior art, or a complete or partial classification of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0181] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

Claims

1. A travel recommendation method, characterized in that, The travel recommendation methods include: Determine a set of candidate points of interest based on the user's travel voice text; Based on the user's historical access points of interest, a set of historical access points of interest is generated; wherein, the historical access points of interest in the set of historical access points of interest include access time characteristics and cabin status characteristics. Based on the candidate point of interest set and the historical point of interest set, a spatiotemporal relationship matrix is ​​determined; wherein, the spatiotemporal relationship matrix is ​​used to characterize the road network detour cost between the first target node and the second target node; the first target node is a node in the candidate point of interest set; the second target node is a node in the historical point of interest set; Based on the spatiotemporal relationship matrix, the set of historical access points of interest, and the set of candidate points of interest, travel points of interest are recommended for the user.

2. The travel recommendation method according to claim 1, characterized in that, The step of recommending travel interest points to the user based on the spatiotemporal relationship matrix, the set of historical access points of interest, and the set of candidate points of interest includes: Determine the first feature matrix corresponding to the set of historical access points of interest; Based on the spatiotemporal relationship matrix, attention learning is performed on the first feature matrix to obtain the second feature matrix. The attention of the target historical visit interest points in the first feature matrix is ​​determined based on the road network detour cost between the target historical visit interest points and the target candidate interest points. The target candidate interest points are candidate interest points in the spatiotemporal relationship matrix that have a road network detour relationship with the target historical visit interest points. Based on the candidate interest point set and the second feature matrix, the target interest point is determined; The target points of interest are recommended to the user.

3. The travel recommendation method according to claim 2, characterized in that, The step of determining the target interest point based on the candidate interest point set and the second feature matrix includes: Based on the travel voice text, determine the travel intention vector; Based on the travel intention vector, attention learning is performed on the second feature matrix to determine the third feature matrix; The target interest point is determined based on the candidate interest point set and the third feature matrix.

4. The travel recommendation method according to claim 3, characterized in that, The step of determining the target interest point based on the candidate interest point set and the third feature matrix includes: Based on the third feature matrix, the travel score of each candidate point of interest in the candidate point of interest set is determined; The predetermined number of candidate points of interest ranked first by the travel score are determined as the target points of interest.

5. The travel recommendation method according to claim 2, characterized in that, Based on the spatiotemporal relation matrix, attention learning is performed on the first feature matrix to obtain the second feature matrix, including: Determine the query matrix, key matrix, and value matrix corresponding to the first feature matrix; Determine the scaled dot product attention matrix of the query matrix and the key matrix; The attention weight matrix is ​​determined based on the fusion matrix of the scaled dot product attention matrix and the fourth feature matrix; wherein the fourth feature matrix is ​​the product of the first feature matrix and the scaling scalar. The second feature matrix is ​​determined based on the attention weight matrix and the value matrix.

6. The travel recommendation method according to claim 1, characterized in that, The road network detour cost includes: the road network detour time cost and the road network detour spatial distance cost.

7. The travel recommendation method according to claim 2, characterized in that, Determining the first feature matrix corresponding to the set of historical access points of interest includes: Determine the basic feature matrix; wherein the basic feature matrix is ​​used to represent the access time characteristics and cabin status characteristics of each historical access point of interest in the historical access point of interest set; Based on the set of historical access points of interest, a first supplementary feature matrix is ​​determined; wherein, the first supplementary feature matrix is ​​used to represent pairs of historical access points of interest with adjacent access times in the set of historical access points of interest. Based on the set of historical access points of interest, a second supplementary feature matrix is ​​determined; wherein, the second supplementary feature matrix is ​​used to represent the difference between the cabin state features of each historical access point of interest in the set of historical access points of interest and the current cabin state features; The first feature matrix is ​​determined based on the basic feature matrix, the first supplementary feature matrix, and the second supplementary feature matrix.

8. The travel recommendation method according to claim 7, characterized in that, Determining the first feature matrix based on the basic feature matrix, the first supplementary feature matrix, and the second supplementary feature matrix includes: Based on the first supplementary feature matrix and the second supplementary feature matrix, the relative dynamic location index of each historical access point of interest is determined; wherein, the index number of the relative dynamic location index is used to indicate the access order of revisiting each historical access point of interest in the set of historical access points of interest. The first feature matrix is ​​determined based on the relative dynamic position index and the basic feature matrix.

9. The travel recommendation method according to any one of claims 1-8, characterized in that, The access time characteristics include the access time and whether the access time is on a weekday; the cabin status characteristics include whether there are children in the cabin and the number of passengers in the cabin.

10. The travel recommendation method according to claim 3, characterized in that, The candidate interest point set and the travel intention vector are output by the large language model after the travel speech text is input into the large language model; The travel recommendation method also includes: Determine the user's recommendation feedback for the target point of interest; Based on the recommended feedback, the speech-text recognition capability of the large language model is optimized.

11. A travel recommendation device, characterized in that, The travel recommendation device includes a determination module, a generation module, and a recommendation module; The determining module is used to determine a set of candidate points of interest based on the user's travel voice text; The generation module is used to generate a set of historical access points of interest based on the user's historical access points of interest; wherein, the historical access points of interest in the set of historical access points of interest include access time features and cabin status features. The determining module is further configured to determine a spatiotemporal relationship matrix based on the candidate point of interest set and the historical access point of interest set; wherein, the spatiotemporal relationship matrix is ​​used to characterize the road network detour cost between the first target node and the second target node; the first target node is a node in the candidate point of interest set; and the second target node is a node in the historical access point of interest set. The recommendation module is used to recommend travel interest points to the user based on the spatiotemporal relationship matrix, the set of historical access interest points, and the set of candidate interest points.

12. A vehicle, characterized in that, The vehicle uses the travel recommendation method as described in any one of claims 1-10 to make travel recommendations.

Images