Big model-based topic recommendation method and device

Abstract

This disclosure provides a method, apparatus, electronic device, storage medium, and computer program product for topic recommendation based on a large-scale model, relating to the field of artificial intelligence technology, specifically large-scale models, intelligent agents, natural language understanding, electronic maps, and point-of-interest (POI) recommendation. It can be applied to topic recommendation scenarios in map applications. The specific implementation scheme is as follows: for an object issuing a non-immediate travel search request, a context sequence is obtained by combining the spatiotemporal context of the object's historical search requests; using an artificial intelligence large-scale model, the non-immediate travel intent set of the object is determined based on the context sequence; using the artificial intelligence large-scale model, the set of recommended topics to be displayed in the object's map application is determined based on the non-immediate travel intent set. This disclosure improves the accuracy of topic recommendation results by recommending non-immediate travel topics in map applications using an artificial intelligence large-scale model, meeting the browsing needs of objects during non-immediate travel phases.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence, specifically to the technical fields of large models, intelligent agents, natural language understanding, electronic maps, and point-of-interest recommendation. In particular, it relates to a topic recommendation method, device, electronic device, storage medium, and computer program product based on a large model, which can be applied to topic recommendation scenarios in map applications. Background Technology

[0002] Currently, the homepages of mainstream map applications only cover the scenario of "immediate / definite travel," assuming that opening the map means the user needs to immediately set off for a destination. However, a large number of users have browsing needs in the "non-immediate travel" stage, including leisurely exploration, travel planning, and obtaining information about the surrounding area. In these scenarios, there is a lack of effective content supply and demand stimulation mechanisms. Summary of the Invention

[0003] This disclosure provides a method, apparatus, electronic device, storage medium, and computer program product for topic recommendation based on a large model.

[0004] According to the first aspect, a topic recommendation method based on a large model is provided, including: for an object that makes a non-real-time travel search request, combining the spatiotemporal context of the object's historical search requests to obtain a context sequence; using an artificial intelligence large model, determining the object's non-real-time travel intention set based on the context sequence; and using an artificial intelligence large model, determining the set of recommended topics to be displayed in the object's map application based on the non-real-time travel intention set.

[0005] According to the second aspect, a topic recommendation device based on a large model is provided, comprising: a sequence generation unit configured to obtain a context sequence for an object that issues a non-immediate travel search request, by combining the spatiotemporal context of the object's historical search requests; an intent determination unit configured to determine the object's non-immediate travel intent set based on the context sequence using an artificial intelligence large model; and a topic determination unit configured to determine the set of recommended topics to be displayed in the object's map application based on the non-immediate travel intent set using an artificial intelligence large model.

[0006] According to a third aspect, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method as described in any implementation of the first aspect.

[0007] According to a fourth aspect, a non-transitory computer-readable storage medium is provided that stores computer instructions for causing a computer to perform the method described in any implementation of the first aspect.

[0008] According to a fifth aspect, a computer program product is provided, comprising: a computer program that, when executed by a processor, implements the method as described in any implementation of the first aspect.

[0009] According to the technology disclosed herein, a topic recommendation method and apparatus based on a large model are provided. For an object that issues a non-immediate travel search request, a context sequence is obtained by combining the spatiotemporal context of the object's historical search requests. The non-immediate travel intent set of the object is determined based on the context sequence using an artificial intelligence large model. The recommended topic set to be displayed in the object's map application is determined based on the non-immediate travel intent set using an artificial intelligence large model. By recommending non-immediate travel topics in the map application with the help of the artificial intelligence large model, the accuracy of topic recommendation results is improved, and the browsing needs of the object during the non-immediate travel stage are met.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is an exemplary system architecture diagram that can be applied to an embodiment of this disclosure; Figure 2 This is a flowchart of an embodiment of the topic recommendation method based on a large model according to the present disclosure; Figure 3 This is a schematic diagram illustrating an application scenario of the topic recommendation method based on a large model according to this embodiment; Figure 4 This is a flowchart of yet another embodiment of the large-model-based topic recommendation method according to this disclosure; Figure 5 This is a structural diagram of one embodiment of the topic recommendation device based on a large model according to the present disclosure; Figure 6 This is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present disclosure. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] The technical solutions disclosed herein involve the collection, storage, use, processing, transmission, provision, and disclosure of various types of information, such as user personal information, in accordance with relevant laws and regulations and do not violate public order and good morals.

[0014] Figure 1 An exemplary architecture 100 is shown that can be applied to the large-model-based topic recommendation method and apparatus disclosed herein.

[0015] like Figure 1 As shown, the system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. The communication connections between terminal devices 101, 102, and 103 form a network topology. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0016] Terminal devices 101, 102, and 103 can be hardware or software that supports network connectivity for data acquisition, interaction, and processing. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices that support network connectivity, information acquisition, interaction, display, and processing functions, including but not limited to smartphones, tablets, e-book readers, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices. They can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are imposed here.

[0017] Server 105 can be a server that provides various services, such as recommending non-real-time travel topics in a map application in response to an object launching a map application via terminal devices 101, 102, or 103, using a large artificial intelligence model. As an example, server 105 could be a cloud server.

[0018] It should be noted that a server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (such as software programs or software modules used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0019] It should also be noted that the topic recommendation method based on a large model provided in the embodiments of this disclosure is generally executed by a server, but the possibility of it being executed by a terminal device, or by a combination of the server and the terminal device, is not excluded. Accordingly, the various parts (e.g., various units) included in the topic recommendation device based on a large model can be all located in the server, all located in the terminal device, or separately located in the server and the terminal device.

[0020] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included. When the electronic devices running the large-model-based topic recommendation method do not require data transmission with other electronic devices, the system architecture may consist only of the electronic devices (e.g., servers or terminal devices) running the large-model-based topic recommendation method.

[0021] Please refer to Figure 2 , Figure 2 A flowchart of a topic recommendation method based on a large model provided in this disclosure embodiment. Flowchart 200 includes the following steps: Step 201: For an object that issues a non-real-time travel search request, combine the spatiotemporal context of the object's historical search requests to obtain a context sequence.

[0022] In this embodiment, the execution entity of the topic recommendation method based on the large model (e.g., Figure 1 For an object that issues a non-real-time travel search request, the server in the middle obtains a context sequence by combining the spatiotemporal context of the object's historical search requests.

[0023] Non-instant travel search requests refer to search behaviors initiated by users in map applications that represent potential future travel planning intentions rather than immediate navigation tasks. These requests are characterized by low time sensitivity, focusing on information gathering, attraction exploration, or itinerary planning. They typically manifest as long-distance queries to specific geographical locations, fuzzy searches within broad tourism categories, or continuous browsing of multiple tourism-related locations within the same city within a short period.

[0024] Non-instant travel search requests are, for example: For example, a user whose permanent residence is A might directly search for "Disneyland in location B" in a map application. This request spans the user's permanent city, and the target POI (Point of Interest) is a typical tourist attraction, indicating that the user has future travel plans to another location, rather than current commuting needs.

[0025] Fuzzy exploratory behavior based on general entertainment intent, such as a user entering "what's fun nearby" or "food guide for location A". Although such requests may be based on the current location, the keywords include general entertainment attributes such as "fun" and "guide", and exclude specific navigation instructions (such as "go to the airport"), which are exploratory searches for reference in future travel decisions.

[0026] Continuous browsing behavior targeting multiple attractions within the same city, such as a user searching for "Summer Palace," "Yuanmingyuan," and "Great Wall" in quick succession, constitutes a strong non-immediate travel signal for in-depth local tours or weekend getaways.

[0027] Spatiotemporal context refers to a structured data set extracted from a user's historical search logs, which can reconstruct the physical and temporal environment in which the user's historical behavior occurred. It is the basic unit for constructing the evolution trajectory of user interests and is used to characterize the specific information needs generated by a user at a specific time and place.

[0028] The spatiotemporal context includes information such as query timestamp, geographic location, and hit POI. A complete spatiotemporal context instance can be represented as a triple structure: {query timestamp, geographic location, hit POI}.

[0029] A query timestamp is a precise record of the exact time a user initiates a search request. Technically, it typically represents the system time when the server receives the client's request, or a timestamp generated and uploaded locally by the client. This data includes not only date and time information (such as year, month, day, hour, minute, and second), but also implicitly contains periodic characteristics of the time dimension (such as weekdays / holidays, day / night, and seasons).

[0030] Geographic location refers to the physical spatial coordinates or administrative region of a user at the moment a search request is initiated. This location information can be obtained through various technical means, including but not limited to GPS signals, base station positioning data, wireless network fingerprint positioning, or based on the user's designated city of residence or the currently manually selected city area. In terms of data structure, it can be represented as latitude and longitude coordinate pairs, geofence markers, or administrative division codes.

[0031] A hit POI refers to the specific point of interest entity data that a map application system matches and displays to the user based on search keywords after the user initiates a search request, or that the user further interacts with. It includes not only the POI's basic name and category attributes (such as attractions, hotels, restaurants), but also its ranking position in the search results list, its display format, and the user's interactive feedback (such as viewing details, saving, sharing). It is the concrete projection of the user's search intent into the physical world.

[0032] As an example, when a non-immediate travel search request is detected from the current object, the current system time is first obtained as a baseline. Based on this baseline, a preset time period (e.g., the past 7 days or 30 days) is traced back to define an initial historical search mining window.

[0033] Within this mining window, all historical search records for the object are traversed. For each record, its query timestamp, the geographical location at which the search was initiated, and the POI information that was hit or clicked by the search are extracted to form the original spatiotemporal context set.

[0034] To filter out weakly relevant noise such as daily commutes, geofencing is established based on the object's permanent residence. If the geographical location in a historical search request is within the permanent residence geofencing, it is further determined whether the matched POI belongs to the "general tourism category" (such as attractions, parks, museums). If it does, the spatiotemporal context is retained; otherwise, it is removed.

[0035] If the geographic location in a historical search request is outside the permanent residence fence (i.e., in a different location), then regardless of the POI hit, it is considered a high-weight potential travel clue, and the spatiotemporal context is directly preserved.

[0036] All valid spatiotemporal contexts after filtering are sorted in ascending order strictly according to the query timestamps. The sorted spatiotemporal contexts are then concatenated to form an ordered list structure, which is the aforementioned context sequence.

[0037] As another example, retrieve the historical search request stream of an object within a preset time period. Use time interval thresholds (e.g., consider two searches with an interval of more than 30 minutes as a new session) to segment continuous search behavior into several independent search session segments.

[0038] For each search session segment, the multiple original search records contained within it are aggregated in the following multi-dimensional way: 1. Time aggregation: Take the timestamp of the first search in the session as the start time of the session, and the timestamp of the last search as the end time.

[0039] 2. Spatial aggregation: Calculate the center point of all search origination locations in this session, and use it as the representative geographical location of this session.

[0040] 3. Content Aggregation: All POIs and search terms involved in the session are aggregated to form the topic tags for the session (e.g., {First POI, Second POI, Flag Raising Ceremony} -> “Tour of Beijing Central Axis”).

[0041] Thus, a conversation segment is transformed into a higher-order spatiotemporal context unit.

[0042] Feature matching is performed on the aggregated spatiotemporal context units. A unit is included in the candidate set only if it meets any of the following conditions: The search terms within the unit contain words with clear travel intent (such as "travel guide", "tickets", "nearby trips").

[0043] The geographical location represented by the unit has changed across cities (e.g., from Beijing to Tianjin).

[0044] If the number of POIs contained in a unit exceeds a preset threshold (e.g., more than 3 attractions are searched in a single session), it indicates a deep planning behavior.

[0045] The selected high-order spatiotemporal context units are linked in chronological order of their start times. The resulting context sequence is no longer a list of single points, but a chain composed of "event blocks" representing specific stages of exploration.

[0046] In some optional implementations of this embodiment, the execution entity performs step 201 as follows: for an object that issues a non-real-time travel search request within the first time period ending now, a context sequence is obtained by combining the spatiotemporal context of the object's historical search requests within the second time period ending now, wherein the duration of the first time period is less than the duration of the second time period.

[0047] First, the system sets a relatively short time window as the first time period. This first time period is usually defined as a recent time range going back from the current moment, such as the past hour.

[0048] The system monitors or periodically scans the behavior logs of all users in real time to retrieve whether at least one non-immediate travel search request has been generated within the first time period up to the present. If an object has search queries with obvious future planning attributes within the first time period, such as "Where to go next summer vacation", "National Day airfare price trend" or "Sanya winter hotels", then the object is marked as an "active non-immediate travel intention candidate".

[0049] For target objects selected in the first time period, the system further initiates the mining of long-term historical data, setting a second time period with a significantly larger span than the first time period. This second time period is typically defined as a longer period looking back from the current moment, such as the past 7 days. Within this time range, the system extracts all historical search requests generated for the object and parses the spatiotemporal context information contained in these requests.

[0050] Finally, the system combines the temporal context obtained in the second time period to generate a context sequence.

[0051] In this implementation, by setting first and second time periods of different durations, the immediacy of intent triggering and the comprehensiveness of user profiles are combined. It can accurately capture current non-immediate travel needs using short time windows, and eliminate accidental search noise by using the historical spatiotemporal context of long time windows, which significantly improves the accuracy and robustness of understanding users' deep travel intentions.

[0052] In some implementations, in response to the detection of a search request data retrieval failure within the first time period ending now, the processing flow of the current batch is skipped; in response to the detection of a search request retrieval failure within the first time period ending now for a preset number of consecutive times, a preset level of alarm notification is triggered.

[0053] The system monitors the status of the data pipeline in real time. If no user's search request data is successfully obtained within the currently set time window (i.e., "the first time period up to the present", such as the past 1 hour), it is determined that the data acquisition has failed (i.e., the data flow is interrupted).

[0054] Once it is confirmed that there is no data input within a certain time period, the system will not perform invalid calculations or idle waiting, but will directly skip the current batch of data processing tasks to avoid wasting computing resources.

[0055] To prevent occasional network fluctuations from being mistaken for serious faults, the system has a fault tolerance threshold. A serious alarm will only be triggered if the search request fails to be retrieved in three consecutive batches (i.e., three consecutive detections), notifying operations and maintenance personnel to intervene and investigate, thereby reducing the false alarm rate while ensuring response speed.

[0056] Step 202: Using a large artificial intelligence model, determine the set of non-real-time travel intentions of the object based on the context sequence.

[0057] In this embodiment, the aforementioned execution entity uses an artificial intelligence big data model to determine the set of non-real-time travel intentions of the object based on the context sequence.

[0058] Large-scale AI models refer to general-purpose AI systems based on deep neural network architectures, pre-trained on massive multimodal corpora, and possessing complex semantic understanding and reasoning generation capabilities. In this solution, the model serves as the core computing engine for intent recognition and topic generation. Its technical feature lies in its ability to receive structured context sequences as input. Through an internal parameterized attention mechanism and a multi-layer Transformer architecture, it performs nonlinear feature extraction and high-dimensional semantic mapping on the input data, ultimately outputting intent labels or recommended topics that conform to human cognitive logic. Examples include large language models and multimodal large language models.

[0059] Non-immediate travel intent refers to the psychological need or decision-making tendency for potential travel planning within a specific future time period, derived from a user's historical search behavior sequence in map applications. Unlike immediate navigation needs, this intent is characterized by a long time span, uncertain decision-making cycle, and discrete behavioral manifestations. It is not a direct mapping of a single search behavior, but rather a high-level semantic concept abstracted from the user's search content over a longer time window through pattern recognition, used to represent the user's persistent travel desire that has not yet been finalized.

[0060] For example, a user might start intensively searching for tourist attractions in a destination (such as Dali, Yunnan) a week ago, then intersperse this with several local life-related queries (such as nearby supermarkets and gas stations) over the next few days, and recently resume searching for transportation or accommodation in that destination. This intermittent search pattern, exhibiting a "start-interruption-restart" over time, is key evidence distinguishing between short-term, impromptu needs (such as finding a restroom) and long-term planning intentions. Despite the intervening noise from local life, large-scale AI models can identify this periodic return of focus spanning several days, thus confirming that the user has a continuous, non-immediate "Dali travel planning" intention.

[0061] As an example, firstly, the acquired context sequence is preprocessed, and historical search requests are sorted according to the order of query timestamps to form an ordered behavior flow. Secondly, a sliding time window mechanism is used to extract multiple sub-sequences from the ordered behavior flow. For each sub-sequence, the textual semantic similarity between each search request is calculated. If the semantic similarity of multiple consecutive search requests in a sub-sequence exceeds a preset threshold, and the distance between the geographical location corresponding to these requests and the user's current residence exceeds a preset spatial threshold, then the sub-sequence is marked as an "out-of-town exploration segment".

[0062] Next, analyze the time intervals between the various "out-of-town exploration segments". If two or more "out-of-town exploration segments" point to the same geographical area or similar tourism themes (e.g., both involve seaside vacations or ancient town visits), and the time interval between them is within a preset non-real-time planning period (e.g., 3 to 30 days), then link the two segments together.

[0063] Finally, the associated fragments are merged into a complete non-immediate travel intent, and then multiple non-immediate travel intents are combined to obtain a set of non-immediate travel intents. For example, the fragments of searching for "Sanya hotels" three days ago and searching for "Sanya flights" today, combined with the interval in between, are determined to be the non-immediate travel intent of "planning to travel to Sanya".

[0064] As another example, we first construct a state transition model that includes various non-real-time travel scenarios. This model defines typical search behavior characteristics corresponding to different stages, from "interest emergence" to "information gathering" and then to "trip confirmation." For example, the "interest emergence" stage is usually characterized by broad destination term searches, the "information gathering" stage is characterized by specific attraction and travel guide searches, and the "trip confirmation" stage is characterized by practical information searches such as transportation and accommodation.

[0065] Secondly, the context sequence is input into a large-scale artificial intelligence model, which maps each search request in the sequence to a node in a state machine. The model not only focuses on the attributes of individual nodes but also analyzes the transition logic between nodes.

[0066] Secondly, the model searches for the path with the highest matching degree to the current context sequence in the state transition model. If the search requests in the context sequence show a trend of progressing from "broad destinations" to "specific attractions" and then to "practical services", the model can still identify this main trajectory even if other irrelevant searches are interspersed in between.

[0067] Finally, based on the status of the matched path's endpoint, the current set of non-real-time travel intentions is determined. For example, if the matched path is still in the "information gathering" stage, the determined intention is "researching travel guides for a certain place"; if the path has progressed to the "trip confirmation" stage, the determined intention is "preparing to book a trip to a certain place".

[0068] In some optional implementations of this embodiment, the above-mentioned execution entity performs the above-mentioned step 202 in the following manner: First step, convert the spatiotemporal quantization data in the spatiotemporal context into spatiotemporal semantic data to obtain a converted data pair including spatiotemporal quantization data and spatiotemporal semantic data; Second step, determine the non-real-time travel intention set of the object based on the context sequence and the converted data pair through the artificial intelligence big model.

[0069] Spatiotemporal quantized data refers to objectively recorded data existing in a spatiotemporal context in the form of numerical values, coordinates, or standardized codes. It is characterized by precision, discreteness, and strong machine readability, but lacks direct human cognitive meaning. Specifically, it includes latitude and longitude coordinate pairs representing geographical locations (such as floating-point numbers in the GCJ02 coordinate system), Unix timestamps or standard date format strings representing points in time, and purely numerical fields representing distance or speed.

[0070] Spatiotemporal semantic data refers to descriptive text or tag data with clear human cognitive meaning, generated from spatiotemporal quantified data through geographic information system parsing, calendar rule matching, or logical deduction. It maps abstract numerical values ​​to specific scene concepts, such as mapping coordinates to administrative divisions or points of interest names, and mapping time points to commuting peaks, holidays, or seasonal characteristics.

[0071] For example, the spatiotemporal quantification data "116.48,39.92" was transformed into the semantic data "Wangjing Business District, Chaoyang District, Beijing", realizing spatial anchoring from coordinates to specific life scenarios; the spatiotemporal quantification data "Friday 18:00" was transformed into the semantic data "Friday Evening Rush Hour / Eve of the Weekend", giving the time point a combined inference meaning of time period-week based on traffic congestion expectations and psychological relaxation; the spatiotemporal quantification data "October 1st" was transformed into the semantic data "National Day Golden Week", clarifying the statutory holiday attribute of this date.

[0072] First, the context sequence of the user in the map application is obtained, and the raw spatiotemporal quantization data is extracted from it. This data is typically represented by GPS (Global Positioning System) location points, millisecond-level timestamps of the search initiation time, etc.

[0073] Secondly, a spatiotemporal semantic transformation operation is performed to generate transformed data pairs. This process includes two parallel branches: 1. Spatial Semanticization Branch: This branch calls a reverse geocoding service or a pre-built geofence database to match the extracted latitude and longitude coordinates with electronic map data. If the coordinates fall within a specific administrative region or well-known landmark area, a corresponding place name entity label is generated (e.g., "Zhongguancun, Haidian District, Beijing"). Simultaneously, it compares the user's current location with their usual residence. If the distance between the current coordinates and their usual residence exceeds a threshold, a status label of "out of town" or "traveling" is generated.

[0074] 2. Time Semanticization Branch: This branch parses timestamp data and performs multiple mappings using a built-in perpetual calendar library and urban traffic pattern models. On one hand, it determines whether the time point falls within a national statutory holiday or a adjusted workday, generating calendar semantics such as "Spring Festival holiday" or "workday." On the other hand, based on the city's historical traffic flow model, it determines whether the time point is during "morning rush hour," "late night," or "off-peak hours," generating time-period behavioral semantics.

[0075] In this way, the original quantitative data (Key) is bound with the generated semantic data (Value) to form a structured transformation data pair.

[0076] Next, the converted data is concatenated or merged with the original context sequence (including user search terms, click behavior, etc.) to construct enhanced input prompts for the large model. At this point, the information input to the large model is no longer just "searched for a hotel at a certain coordinate on October 1st", but becomes "during the National Day Golden Week, the user was in a different location and searched for a hotel".

[0077] Finally, the enhanced input information is processed by a large AI model. This model utilizes its pre-trained knowledge or web search instructions to focus on analyzing the implicit scenario constraints within the spatiotemporal semantic data. For example, when it identifies the coexistence of the semantic tags "Friday evening rush hour" and "out-of-town," the model infers that the user may be planning an upcoming weekend trip or returning from a business trip, thus filtering and identifying a set of non-immediate travel intentions that fit the current spatiotemporal context from numerous possible intentions.

[0078] In this implementation, by transforming abstract spatiotemporal quantitative data into spatiotemporal semantic data rich in scene information, the ambiguity of numerical data in semantic understanding is eliminated. This enables large artificial intelligence models to accurately capture users' potential non-immediate travel needs by combining deep background information such as holidays and geographical locations, significantly improving the accuracy and intelligence level of intent recognition.

[0079] In some optional implementations of this embodiment, the execution entity performs the second step described above in the following manner to determine the set of non-immediate travel intentions of the object: by using an artificial intelligence big data model, based on the context sequence, conversion data pairs, the object's profile data, and prior guidance data, to determine the set of non-immediate travel intentions of the object.

[0080] Profile data refers to a structured set of labels or statistical feature vectors that characterize an object's long-term behavioral characteristics, preferences, and social attributes. Unlike contextual sequences that reflect instantaneous states, profile data possesses relative stability and cumulativeity, and is used to depict the "static" or "quasi-static" attributes of an object. As a background constraint for intent reasoning, it can assist large-scale artificial intelligence models in performing personalized intent completion or correction based on the object's historical habits when information is sparse or ambiguous.

[0081] Prior guidance data refers to external knowledge constraints injected into the model during inference to standardize the output format or limit the direction of inference. It does not contain the current user's real-time state, but rather represents pre-set logical rules, recommendation strategies, or standard examples from the business side. Its purpose is to converge the capabilities of general-purpose AI models to specific business scenarios, preventing the model from generating illusions or outputting intent labels that do not conform to the product definition.

[0082] Prior guidance data may include, for example: Directional reference, for example: "You are a professional travel planning assistant. Please select intent tags only from the following predefined list: [Nearby trips, long-distance travel, local life exploration, commuting navigation]. If the user's intent is unclear, prioritize categorizing it as 'local life exploration'." Few-Shot example, such as "Friday evening rush hour, search Wangjing business district" -> Expected output: {Intent: commuting home, urgency: high}.

[0083] As an example, firstly, the long-term attribute tags of the object (such as the city of residence, historically frequent travel destinations, consumption level, family member structure, etc.) are read from the user database and serialized into natural language descriptions or structured JSON (JavaScript Object Notation) strings to form user profile text blocks.

[0084] Secondly, construct a prompt template containing prior guidance data. The prompt template consists of three parts: Task Instruction Area: Clearly inform the AI ​​model that the current task is to "identify non-immediate travel intentions" and list all legal intention categories (i.e., a finite set defined by the business side).

[0085] Sample Demonstration Area (Few-Shot): Select 3-5 typical mapping cases from "spatiotemporal semantic data + search terms" to "standard intent" and embed them into the inference demonstration template.

[0086] Constraints section: Set inference rules, such as "When the user profile data shows that the user is price-sensitive, if the search term contains high-end hotels, it should be judged as 'searching for others' or 'specific needs' rather than 'personal regular travel'."

[0087] Next, multi-source data concatenation is performed. The real-time acquired context sequence, conversion data pairs, user profile text blocks, and the aforementioned "prompt template" are concatenated into a complete input Prompt in a preset order. During this process, an attention masking mechanism is used to ensure that the model can simultaneously focus on real-time spatiotemporal semantics and long-term profile features when generating intent.

[0088] Finally, a large-scale AI model is invoked for inference. The model receives the concatenated complete input, processes it through a self-attention mechanism, and outputs intent labels and their confidence levels that conform to the prior guidance data specifications. The system receives the output results and encapsulates them into a "set of non-real-time travel intents." If the intent output by the model is not in the predefined list, it is corrected or discarded according to the fallback strategy in the prior guidance data.

[0089] In this implementation, the introduction of profile data enables personalized adaptation of intent recognition, resolving the ambiguity of the same search terms varying from person to person; combined with prior guidance data, the generation boundary of the large model is effectively constrained, ensuring that the output intent conforms to business specifications, and significantly improving the accuracy and usability of the inference results.

[0090] In some optional implementations of this embodiment, the execution entity determines the set of non-immediate travel intentions of the object in the following manner: First, based on the context sequence, conversion data pairs, profile data, and prior guidance data, it determines the object's non-immediate travel intentions and at least one of the following: the clarity, priority, and failure condition of the non-immediate travel intentions; then, by combining each non-immediate travel intention and at least one of the following: clarity, priority, and failure condition, it obtains the set of non-immediate travel intentions. Here, "at least one" can be any one, two, multiple, or even all of them.

[0091] Explicitness refers to the quantitative representation of the semantic certainty or confidence level of a large artificial intelligence model regarding non-immediate travel intentions. It reflects whether the user's current search behavior expresses a specific need or is merely in a vague exploratory stage. Explicitness is usually expressed in numerical or discrete label form to distinguish the clarity levels of intent.

[0092] For example, under high explicitness, a user searches for "book a high-speed rail ticket from Beijing to Shanghai on May 1, 2024". This request includes specific time, location, mode of transportation, and actions, and the intent is clear and unambiguous. The explicitness can be marked as "Explicit" or with a confidence level of 0.95.

[0093] In a low-clarity search, a user searches for "where to go for the May Day holiday". This request only contains a time constraint and lacks a destination and specific activity, which is a typical example of fuzzy exploration. The clarity can be labeled as "Fuzzy" or have a confidence level of 0.4.

[0094] Priority refers to the ranking of the display or processing weight of each intent relative to other intents within the same set of non-immediate travel intents. It comprehensively reflects the urgency of the intent, the degree of matching with the user profile, and the fit with the current time and space scenario, and is used to determine which intents should be displayed or processed first within the limited interface space.

[0095] Suppose a user has two non-immediate travel intentions: "weekend camping" and "long-distance trip during the National Day holiday." If the current time is Thursday afternoon and the user is located in the suburbs, "weekend camping" is prioritized as "high" due to its proximity and high relevance to the scenario; while "long-distance trip during the National Day holiday," although a clear intention, is prioritized as "medium" because it is further away. During the display, the system will prioritize showing recommended topics related to "weekend camping."

[0096] Failure conditions refer to the triggering rules embedded in non-immediate travel intentions, used to determine whether the intention should be terminated or exited from the processing flow. As metadata for intention lifecycle management, it enables the system to automatically identify the completion, expiration, or replacement status of an intention, thereby promptly cleaning up invalid recommendations and avoiding resource waste and user experience interference.

[0097] For example, failure conditions include: 1. Intent Completed: If all query requests associated with an intent have reached the final state of "navigated", "reached destination" or "expired", then the intent is deemed to have met the invalidation condition and will automatically exit.

[0098] 2. Expired Intent: If the latest related query request under this intent is more than 7 days ago, and there have been no new related search behaviors during this period, then the intent is determined to have failed due to loss of timeliness.

[0099] 3. Intent overridden by new intent: If the AI ​​model marks the superseded_by field in the output and points to a new intent ID (Identification), then the old intent is deemed to have met the invalidation condition and is taken over by the new intent for display.

[0100] As an example, first, a composite input prompt is constructed, containing context sequences, transformation data pairs, profile data, and prior guidance data. In the prompt template, the output structure is explicitly defined as a JSON object containing the intent key field and a JSON object containing at least one of the following key fields: clarity, priority, and exit_condition.

[0101] Secondly, a large-scale artificial intelligence model is invoked for joint reasoning. While generating the intended content, the model evaluates the input data based on at least one of the following dimensions using its internal attention mechanism: By analyzing the semantic specificity and contextual coherence of search terms, we evaluate and output precision labels; By comparing the matching strength between the intent content and the current spatiotemporal semantics and user profile, a priority score is evaluated and output. By identifying the completion status, time span, and semantic substitution relationship with new intentions, corresponding failure condition meta-information is generated.

[0102] Next, the system receives the structured results from the large model. The system parses and validates the output to ensure that multiple fields conform to the preset data format and value range. If the failure condition in the model output is "overridden by a new intent," the system simultaneously extracts the alternative intent ID from the `superseded_by` field to establish the association between intents.

[0103] Finally, the parsed intent content, along with at least one of the parsed clarity, priority, and expiration conditions, is encapsulated into a unified intent object. All intent objects are then combined into a non-instant travel intent set. This set not only contains "what the user wants to do," but also fully records "how certain this intent is," "where it should rank," and "when it should disappear," providing a complete data foundation for subsequent sorting, display, and lifecycle management.

[0104] In this implementation, the intent, its explicitness, priority, and failure conditions are output synchronously in a single inference, thus achieving the integration of intent recognition and attribute evaluation, avoiding information loss and delay in multi-step processing; the embedded failure condition meta-information enables the intent to have self-management capabilities, significantly improving the real-time response efficiency and resource utilization of the recommendation system.

[0105] In some optional implementations of this embodiment, priority is determined as follows: using a large-scale artificial intelligence model, priority is determined based on at least one of the following: the timeliness and clarity of the non-immediate travel intention, the feedback data of the object to the historical recommended topics corresponding to the non-immediate travel intention, and the fit between the non-immediate travel intention and the current spatiotemporal scenario. Here, "at least one" can be any one, two, multiple, or even all of these factors.

[0106] The timeliness of non-immediate travel intentions refers to a quantitative indicator representing the urgency or novelty of that intention relative to the current moment. Its calculation logic is typically based on normalization of the maximum time difference between historical query records associated with the intention, following the principle of "the closer to the present, the higher the priority." It reflects the immediate activity status of user needs and is used to distinguish between long-term planning and short-term impulses.

[0107] The feedback data of the object to the historical recommended topics corresponding to the non-immediate travel intention refers to the statistical characteristics of the user's actual response behavior to similar or related intention recommended content in past interactions. Its calculation logic adopts a tiered weighting coefficient strategy, that is, assigning different scores according to the positiveness of the feedback: clicked behavior > no clicked behavior > displayed but not clicked.

[0108] Suppose the system previously recommended the "Family Fun Park" theme to a user (corresponding to the "Traveling with Kids" intent), and the user clicked and viewed the details, then the feedback data for that intent is recorded as a high score (e.g., 1.0). If the system recommended the "Popular Instagrammable Spots" theme, and the user only saw it but did not click, it is recorded as a medium score (e.g., 0.5). If the system recommended the "Hiking" theme, and the user swiped away directly or did not generate any exposure or clicks, it is recorded as a low score (e.g., 0.1). This feedback data will directly participate in the weighted calculation of priority.

[0109] The fit between non-immediate travel intentions and the current spatiotemporal context refers to a semantic similarity metric that measures the degree of match between a specific intention and the user's current time, location, and environmental context. It relies on the spatiotemporal semantic data generated in previous steps to determine whether the current physical environment constitutes an ideal opportunity to trigger the intention.

[0110] For example, a highly compatible scenario is as follows: The current spatiotemporal semantics are "Friday evening rush hour" and the user is located in the "office area". At this time, the non-immediate travel intention is "short trip to the surrounding area" or "weekend relaxation". The two are highly complementary in semantics (the need to release work pressure), and the compatibility score is high.

[0111] Low fit scenarios are as follows: The current spatiotemporal semantics are "weekday morning rush hour" and the user is in a "subway station". If there is an intention to "deep cultural tour", although the intention exists, it conflicts with the current hurried commuting scenario, and the fit score is low.

[0112] First, for each intent instance in the set of non-immediate travel intents, an evaluation vector with at least one of the following dimensions is extracted and constructed.

[0113] 1. Timeliness Dimension Calculation: Obtain the timestamps of all related query records under this intent, and calculate the difference between the maximum timestamp and the current system time. Use a preset decay function (such as exponential decay or linear normalization) to convert this time difference into a timeliness score between 0 and 1. The smaller the time difference, the closer the score is to 1.

[0114] 2. Obtaining the Explicitness Dimension: Directly use the explicitness values ​​determined in the previous steps. For example, assign a value of 1.0 to explicit intent and a value of 0.5 to implicit intent as the base confidence weight.

[0115] 3. Feedback Data Dimension Mapping: Retrieve historical interaction logs for this intent category from the user profile database. If a "click" record exists, set the mapping coefficient to 1.0; if only an "impression" record exists but no clicks, set the mapping coefficient to 0.6; if there are no interaction records, set the mapping coefficient to 0.3. This step aims to incorporate real user behavior to calibrate the model's predictions.

[0116] 4. Fit Determination: Match the current spatiotemporal semantic tags (such as "rainy day", "late night", "business district") with the semantic tags of the intent. If there is a strong correlation rule between the two (such as "rainy day" matching "indoor exhibition"), a high fit score is assigned; if there is a mutually exclusive rule (such as "late night" matching "outdoor hiking"), a very low score is assigned or the score is set to zero.

[0117] Secondly, when only a single-dimensional evaluation vector is obtained, the priority score is determined based on that vector. When multiple-dimensional evaluation vectors are obtained, the constructed multi-dimensional evaluation vectors are input into the AI ​​large-scale model. The AI ​​large-scale model does not simply perform a linear weighted sum; instead, it acts as an inference engine, dynamically adjusting each factor based on pre-trained travel decision logic. For example, when the timeliness is extremely high (the user just searched) but the relevance is extremely low (the current environment does not support it at all, such as searching for travel while in a meeting), the model may lower the immediate display priority of that intent, delaying its processing. Conversely, if the feedback data is excellent (the user has always wanted to see it) and the current time and space scenario perfectly matches (it's finally the weekend), the model will significantly increase its priority.

[0118] Finally, the AI ​​model outputs a comprehensive priority score for each non-immediate travel intention. The system then sorts the set of non-immediate travel intentions in descending order based on this score. For intentions with the same score, a secondary sort is performed based on clarity, prioritizing the display of more clearly defined intentions. The resulting ordered list constitutes the set of non-immediate travel intentions with determined priorities, which will serve as the foundational input data for subsequent steps (such as generating recommendation cards and route planning).

[0119] This implementation introduces multi-dimensional dynamic factors such as timeliness, feedback data, and scenario suitability, and uses a large artificial intelligence model for comprehensive reasoning. This effectively solves the problems of rigid intent ranking and neglect of real-time user status and historical behavior feedback in traditional recommendation algorithms. It can accurately capture the dynamic changes in user needs, ensure that high-value and highly relevant travel intentions are prioritized for identification and display, and significantly improve the accuracy of non-real-time travel intention prediction and the personalization level of user experience.

[0120] Step 203: Using a large artificial intelligence model, determine the set of recommended topics to be displayed in the object's map application based on the set of non-real-time travel intentions.

[0121] In this embodiment, the aforementioned execution entity uses a large artificial intelligence model to determine the set of recommended topics to be displayed in the object's map application based on the set of non-real-time travel intentions.

[0122] Recommended topics refer to content aggregation units or function entry points presented in the user interface of a map application to guide users to the next interaction. These recommended topics are not simply a list of search results, but rather high-level semantic concepts mapped from a set of non-immediate travel intentions. Their forms of expression include, but are not limited to, text titles, icon labels, or interactive cards. Technically, recommended topics act as a bridge between users' vague intentions and specific service resources, aiming to transform abstract "non-immediate travel intentions" into visual, clickable decision options, thereby shortening the path from generating an idea to obtaining specific information.

[0123] For example, when the defined set of non-immediate travel intentions includes "planning a beach vacation," the map application interface can display recommended topics such as "Sanya 5-day travel guide" or "comparison of popular island hotel prices." These two recommended topics correspond to two different sub-categories: "itinerary planning" and "accommodation booking." After a user clicks on the recommended topic "Sanya 5-day travel guide," the map application will display an aggregated page containing attraction routes and food recommendations; while clicking on "comparison of popular island hotel prices" will directly display a list of hotels and price trends.

[0124] As an example, firstly, natural language understanding analysis is performed on the set of non-real-time travel intentions to extract core semantic elements. These core semantic elements include at least the destination entity, travel attributes (such as "food," "history," and "family"), and time constraints (such as "May Day holiday"). Secondly, a pre-built topic generation template library is accessed. This library stores several standard recommendation topic sentence structures, such as "{destination} + {travel attribute} + recommendation" and "{time constraint} + go to {destination} + how to play," etc. Thirdly, the extracted core semantic elements are filled into the recommendation topic sentence structure with the highest matching degree. For example, if the extracted elements are the destination "Xi'an" and the attribute "food," the model will select the template "{destination} + {travel attribute} + must-eat list" to generate the recommendation topic "Xi'an Must-Eat List."

[0125] Finally, the generated recommended themes are encapsulated into UI (User Interface) component data that can be rendered by the map application and sent to the client for display.

[0126] As another example, firstly, the set of non-immediate travel intentions is transformed into input prompts in the form of natural language descriptions. For example, the intention set is transformed into descriptive text such as "The user plans to take their child to Beijing next month and is interested in the science museum and zoo." Secondly, the text generation interface of the large-scale artificial intelligence model is called, with the generation goal set as "generating a map application recommendation title that will attract user clicks." Based on its pre-trained massive knowledge base, the model performs deep reasoning on the input descriptive text, capturing the implicit potential needs such as "family travel," "science popularization," and "trip planning." Thirdly, the model outputs several candidate recommendation topic texts. For example, the model may generate recommendation topics with a natural language style and emotional connotation, such as "Beijing Family Science Popularization Trip: Exploring the Science Museum and Zoo with Children" or "Going to Beijing next month? Please keep this family travel itinerary."

[0127] Finally, target recommended topics that meet the preset length limit and compliance requirements are selected from the candidate recommended topic texts and displayed in the map application.

[0128] In some optional implementations of this embodiment, the execution entity may perform step 203 as follows: using a large artificial intelligence model, determine a set of recommended topics based on prior information about the points of interest corresponding to each non-real-time travel intention in the non-real-time travel intention set.

[0129] Prior information about points of interest (POIs) refers to static or quasi-static descriptive information pre-stored in geographic information databases or knowledge graphs to characterize the inherent attributes, service content, and scene features of a specific POI. This data exists independently of the user's real-time search behavior and serves as a background knowledge base, providing content materials and factual basis for large-scale artificial intelligence models to transform abstract travel intentions into concrete recommendation topics.

[0130] For example, for the point of interest "Shanghai Disneyland", the prior information includes {Category: Theme Park, Tags: [Family, Fireworks Show, Queues], Opening Hours: 09:00-20:00, Suggested Visit Duration: 1 day}.

[0131] As an example, firstly, each non-immediate travel intention in the set of non-immediate travel intentions is traversed. Based on the semantic tags of the intention (such as "family trip", "food exploration", "long-distance travel"), one or more candidate points of interest matching it are retrieved from a pre-built geographic information database. Subsequently, prior information data of these candidate points of interest is extracted to form "intention-point of interest data pairs". For example, for the "food exploration" intention, multiple candidate points of interest are retrieved, and their corresponding categories, tags, average spending per person, and other prior information data are obtained.

[0132] Secondly, using the set of non-real-time travel intentions and the retrieved "intent-point of interest data pairs" as core inputs, prompt words for topic generation are constructed. These prompt words not only contain the user's intent description but also explicitly inject prior information about relevant points of interest. The structure of the prompt words can be designed as follows: "The user currently has a need for [intent name], and relevant location information includes [prior information about point of interest A] and [prior information about point of interest B]. Please generate recommended topics that attract the user based on this information."

[0133] Next, the AI ​​big data model is invoked to process the aforementioned prompts. In this process, the big data model acts as a content curator; it doesn't simply list interest points, but rather performs the following operations: Information filtering and fusion: From a large amount of prior information, the model filters out the key attributes that best match the current non-immediate travel intention. For example, for the intention of "weekend relaxation", the model will focus on extracting the tags "late-night snack" and "local flavor" from the prior data of "Wangjing Xiaoyao", while ignoring non-core information such as "average consumption per person".

[0134] Natural Language Reorganization: This involves semantically fusing the selected intents with interest attributes to generate fluent and engaging natural language phrases as recommended topics. For example, combining the intent of "weekend relaxation" with the attributes of "late-night snacks" and "local flavor" generates the recommended topic "Midnight Diners: Experiencing the Local Flavor of Wangjing".

[0135] Finally, the AI ​​model generates at least one recommended topic for each non-immediate travel intention. The system aggregates all generated recommended topics to form a recommended topic set. Each element in this set is a complete and semantically clear phrase or sentence, directly displayed to the user on the map application interface, thus completing the transformation from "user intention" to "recommended content".

[0136] In this embodiment, by combining non-real-time travel intentions with prior information about points of interest, and leveraging the generative capabilities of large-scale artificial intelligence models, automated and personalized content production from abstract intentions to concrete recommendation topics is achieved. This effectively solves the problems of rigid content templates and low matching degree with user intentions in traditional recommendation systems, significantly improving the attractiveness, relevance, and information richness of recommended content, thereby better stimulating users' travel decisions.

[0137] In some optional implementations of this embodiment, the aforementioned execution entity can determine the recommended topic set in the following manner: First, using an artificial intelligence big data model, multiple recommended topics are determined based on prior information about the points of interest corresponding to each non-immediate travel intention in the non-immediate travel intention set; then, based on at least one of the following: the matching degree between the recommended topic and the decision-making stage of the object, the fit between the recommended topic and the current spatiotemporal scene, and the coverage breadth of the recommended topic, the multiple recommended topics are sorted to obtain the recommended topic set. Here, "at least one" can refer to any one, two, multiple, or even all of them.

[0138] The decision-making stage refers to a classification index that characterizes the cognitive depth and behavioral intention evolution of a user during the travel planning process. It reflects the dynamic psychological process from generating vague ideas to finally determining the itinerary, and is typically divided into multiple linear or progressive levels according to a timeline or cognitive depth. This guides the system to provide differentiated information assistance at different stages. For example, it may include the following stages in sequence: Understanding Phase: Users are in the early stages of need, such as searching "where to go for fun during the May Day holiday". At this time, focus on the macro overview.

[0139] Interest Arousal (Seeding) Stage: Users begin to show interest in specific destinations or activities, such as searching for "Sanya Atlantis Hotel Review" or "Sichuan West Self-Driving Route Recommendation".

[0140] Comparison phase: Users hesitate among multiple alternatives, such as searching "Which is more fun, Disney or Universal Studios?" or "Chengdu vs. Chongqing travel comparison".

[0141] Route planning stage: Users have determined their destination and begin to refine their itinerary, such as searching for "detailed three-day tour guide of Beijing" or "food map around the Forbidden City".

[0142] Future travel phase: Users will conduct long-term, pre-emptive searches, such as "European visa application for next summer" or "skiresort booking in 2025".

[0143] The matching degree between recommended topics and the decision-making stage is used to measure how well the content attributes of the currently generated recommended topics fit the user's current decision-making stage. Its core logic is "supply on demand": in the early stages of decision-making (such as the understanding stage), a high matching degree corresponds to overview-type topics with broad coverage; in the later stages of decision-making (such as the route planning stage), a high matching degree corresponds to tool-type or checklist-type topics with rich details and strong practicality.

[0144] The fit between the recommended topic and the current spatiotemporal context refers to the degree of consistency between the temporal sensitivity and spatial relevance features of the recommended topic and the user's current real-time environmental context (including the current time, geographical location, weather conditions, etc.). It ensures that the recommended content is not only semantically relevant, but also has immediate feasibility or strong relevance in physical time and space.

[0145] The breadth of coverage for recommended topics refers to the number of points of interest, the range of service categories, or the richness of information dimensions covered by the recommended topic. It is used to distinguish whether a topic is macro-level aggregated content (such as "must-eat list of a city" or "top ten attractions") or micro-level single-point content (such as "review of a restaurant" or "tickets for a certain attraction"). Generally, the broader the coverage of a topic, the higher its information density, which is suitable for exploratory scenarios.

[0146] As an example, firstly, for each generated recommended topic, the system calculates a normalized score for at least one of the following dimensions: 1. Decision-making stage matching score: The system identifies the user's current decision-making stage (e.g., "planting the seed" stage). If the recommended topic is "A collection of niche campsites around Chengdu", it is given a high score (e.g., 0.9) because it matches the "planting the seed" characteristic; if the recommended topic is "Chengdu traffic route map", it belongs to the "route planning" stage and does not match the current "planting the seed" stage, so it is given a low score (e.g., 0.2).

[0147] 2. Time and Space Fit Score: Obtain the current time (e.g., 6:00 PM on Friday) and location (e.g., the user is located in a downtown business district). If the recommended theme is "Weekend Suburban Road Trip," it receives a high score because the time is close and meets weekend expectations; if the recommended theme is "Late Night Food Exploration," although the time is suitable, it may be limited by distance, receiving a medium score; if the recommended theme is "Early Morning Flag Raising," it conflicts with the current evening scene, receiving a very low score.

[0148] 3. Coverage Breadth Score: This score analyzes the number of Points of Interest (POIs) associated with the recommended topic. If the topic involves "Top 10 Hot Pot Restaurants in the City," it has many associated POIs and broad coverage, thus receiving a basic weighted score. Conversely, if the topic only involves "a certain hot pot restaurant," its coverage is narrow, resulting in a lower score. This dimension prioritizes displaying aggregated topics with greater information content, unless the user has a very specific, single-point search intent.

[0149] Secondly, when calculating the normalized score for only one dimension, multiple recommended topics are ranked according to the normalized score for that dimension; when calculating the normalized scores for multiple dimensions, multiple recommended topics are ranked according to the weighted fusion of the normalized scores for multiple dimensions.

[0150] Taking the calculation of the normalized scores for the three dimensions mentioned above as an example, a pre-defined ranking algorithm model is used to weight and fuse the scores of the three dimensions. The formula is as follows: Score=w1×Sstage+w2×Scontext+w3×Sbreadth Among them, w1, w2, and w3 are weight coefficients set based on business experience, and Sstage, Scontext, and Sbreadth are the matching score for the decision-making stage, the spatiotemporal fit score, and the coverage breadth score, respectively.

[0151] Finally, based on the calculated overall score, the multiple recommended topics are sorted in descending order to obtain a set of recommended topics. In some examples, a predetermined number of topics (e.g., Top 5) can be selected and assembled into the final set of recommended topics in order.

[0152] This implementation achieves precise alignment between recommended content and user decision-making progress and the real-world environment through multi-dimensional joint sorting; it effectively solves the problem of irrelevant recommendations caused by ambiguous user needs in non-real-time travel scenarios, and improves information acquisition efficiency and user experience.

[0153] In some optional implementations of this embodiment, the execution entity may also perform the following operations: in response to a timeout in the request to call the large artificial intelligence model, perform a retry operation for a preset number of times according to the exponential backoff algorithm; in response to the failure to obtain a valid response after the retry, mark the current task as a failure and add the current task back to the processing queue of the next batch.

[0154] When the system issues a call request but does not receive a response within a specified time (due to network congestion or service overload), it will not wait indefinitely. The system will initiate a retry mechanism, and the retry interval is not fixed, but rather adopts an "exponential backoff" strategy (e.g., a shorter wait time for the first retry, doubling the wait time for the second retry) to reduce server load and improve the retry success rate. The example in the diagram shows two retries.

[0155] If the system still cannot obtain a result after a preset number of retries (e.g., 2), the system will determine that the call has completely failed. At this time, in order to ensure the smooth operation of the overall data processing pipeline, the system will mark the specific task as "failed" and temporarily remove it from the current batch. It will then be automatically retried when the next hourly cycle begins, thus avoiding the impact of a single task blocking the overall progress.

[0156] In some optional implementations of this embodiment, the above-mentioned execution entity may also perform the following operations: in response to the abnormal format parsing of the data returned by the large artificial intelligence model, the randomness in the generation parameters of the large artificial intelligence model is set to zero and a single retry operation is performed; in response to the abnormal format parsing after the retry, the preset default fallback data is called for display.

[0157] Large-scale artificial intelligence models typically exhibit a degree of randomness (Temperature > 0). When the returned data does not conform to the preset JSON format specifications, the system first minimizes the model's "imagination" (i.e., sets temperature=0), causing it to regenerate the content in the most conservative and deterministic manner in an attempt to correct the formatting errors.

[0158] If the content generated in the most conservative mode still fails the format validation, the system will no longer rely on the output of the large AI model, but will directly call a pre-configured general template or static data (as a fallback by default) to fill the gap. This ensures that even if the model completely fails, the user can still see the basic card content, avoiding blank screens or code errors.

[0159] In some optional implementations of this embodiment, the execution entity may also perform the following operation: in response to the processing time of the current batch exceeding a preset time threshold, reduce the number of display cards pushed in the current batch so as to complete the closing of the push window before the preset target time point.

[0160] The system has set strict deadlines for processing tasks (e.g., all tasks must be completed before the Friday evening rush hour). If the system detects that the processing speed of the current batch is too slow and it is expected that all tasks cannot be completed within the specified time (e.g., a 90-minute window), it is considered a timeout.

[0161] To uphold the core objective of "on-time delivery," the system automatically triggers a degradation strategy. Based on a priority algorithm, it prioritizes high-value, high-priority display cards for push notifications, while proactively discarding or delaying the processing of low-priority peripheral messages. This "reduced push" sacrifices some coverage but ensures a core user experience, guaranteeing that most high-quality content is displayed on time within the specified timeframe.

[0162] In some optional implementations of this embodiment, the execution entity may also perform the following operations: in response to failure to push the display card to the preset display terminal, perform a preset number of retry operations; in response to failure to push after retry, write the display card into the dead letter queue to wait for manual compensation processing.

[0163] When pushing the generated card content to the user's preset display terminal, if the sending fails due to network jitter or terminal offline, the system will not immediately discard the message, but will retry multiple times (e.g., 3 times) to make every effort to ensure that the message is delivered.

[0164] If the push fails after multiple retries, it indicates a potential serious link failure or account anomaly. In this case, the system moves the message to a "dead letter queue" for persistent storage. These messages are not automatically deleted but are retained as abnormal data, awaiting manual resending by technical personnel using backend tools. This ensures that every generated recommendation is traceable and prevents complete data loss.

[0165] See also Figure 3 , Figure 3 This is a schematic diagram of application scenario 300 of the topic recommendation method based on the large model in this embodiment. In the past hour, user 301 sent a non-real-time travel search request for "hot pot restaurants near Wangjing" to the map application in server 303 via terminal device 302. For user 301 who sent the non-real-time travel search request, server 301 first combines the spatiotemporal context of user 301's historical search requests over the past 7 days to obtain a context sequence. Then, through the artificial intelligence large model, it determines the set of non-real-time travel intentions of the object based on the context sequence, including the following intentions: (1) "Shanghai travel plan for the National Day holiday" (explicit,priority_score=95); (2) "Looking for food in Wangjing tonight" (fuzzy,priority_score=72).

[0166] Finally, using a large AI model, the set of recommended topics to be displayed in the object's map application is determined based on the set of non-real-time travel intentions, including the following recommended topics: (1) The theme of "Must-visit attractions in Shanghai"; (2) "Quick Selection of Popular Restaurants in Wangjing" theme.

[0167] In this embodiment, for an object that issues a non-immediate travel search request, a context sequence is obtained by combining the spatiotemporal context of the object's historical search requests. Using a large-scale artificial intelligence model, the set of non-immediate travel intentions of the object is determined based on the context sequence. Then, using the large-scale artificial intelligence model, a set of recommended topics to be displayed in the object's map application is determined based on the set of non-immediate travel intentions. By using the large-scale artificial intelligence model to recommend non-immediate travel topics in the map application, the accuracy of topic recommendation results is improved, meeting the browsing needs of the object during the non-immediate travel phase.

[0168] In some optional implementations of this embodiment, the execution entity may also perform the following operations: First, for recommended topics in the recommended topic set, determine the target card template to which the recommended topic belongs from the card template set based on the content volume information and image material information of the recommended topic; then, generate a display card based on the target card template and at least one recommended topic belonging to the target card template; finally, display the display card on the homepage of the map application through a preset display terminal.

[0169] Content volume information refers to attribute parameters used to quantify the richness and structural complexity of the text content contained in the recommended topic. Specifically, this information includes, but is not limited to, the length of the title of the recommended topic, the line limit for the subtitle or introductory text, and the range of related extended query terms. It is a key criterion for determining whether the recommended topic has sufficient information density to support a specific visual style.

[0170] Image resource information refers to the availability and specifications of visual media resources associated with the recommended theme. Specifically, this includes the existence of high-quality cover images, image resolution, aspect ratio, and number of images. This information directly determines whether the display card can use a mixed text and image layout or whether it needs to be downgraded to plain text display.

[0171] Card templates are standardized UI layout containers pre-installed in the front end or configuration center of map applications. They define the visual structure, element arrangement rules, and content filling constraints for displaying cards. Each card template corresponds to a specific content display format. The system dynamically matches the most suitable template type based on the content volume and image material information of the recommended theme to achieve unity between content and form.

[0172] For example, the card template collection includes the following three typical forms: A text search item aggregation template is a card layout container that primarily displays a high-density text list. It is suitable for scenarios with relatively small content (no long titles) and no image materials. Its core element constraints are: containing 6-10 query capsules, with each query length ≤ 12 characters, arranging 3-5 queries per row, and requiring labels such as "Just Searched" or "Repeatedly Searched".

[0173] The multi-image visual media guidance template is an immersive card layout container that uses multiple high-resolution images as the core visual focus. It is suitable for scenarios with abundant image materials (≥2 images) and concise titles (≤16 characters). Its layout features 2-3 large images at the top, accompanied by a single topic hook and 3-5 extended search terms. If there are fewer than 2 images, this template will not be used.

[0174] A single-thumbnail index template refers to a standardized card layout container that uses a horizontal or vertically compact arrangement of text and images, suitable for scenarios with only a single thumbnail (Thumbnail = 1 image). Its layout features a small image displayed on the left or top, accompanied by two lines of topic prompts and 3-5 extended search terms. This template will not be used if there is no thumbnail.

[0175] First, iterate through each recommended topic in the recommended topic set and parse its metadata to obtain information on content volume and image materials.

[0176] If the number of valid high-definition images associated with the recommended topic is greater than or equal to 2, and the title length meets the preset threshold (e.g., ≤16 characters), then it is determined that it meets the hard constraint of multiple visual media guidance templates and is identified as the target card template.

[0177] If it is detected that the recommended topic is associated with only one thumbnail and has the requirement of multiple lines of topic guidance, it is determined that it meets the hard constraint of the single thumbnail index template and is identified as the target card template.

[0178] If the recommended topic lacks effective image materials but contains multiple short text query terms (such as 6 to 10), it is determined that it meets the hard constraints of the text search term aggregation template and is identified as the target card template.

[0179] Next, the specific content of the recommended topic (such as title text, image URL, and extended query list) is populated into the placeholders corresponding to the template. During this process, the system will strictly perform hard constraint validation of the template (for example, if a large image template is selected but only one image is actually retrieved, it will automatically downgrade or rematch), and finally generate a display card data structure that conforms to the UI specifications.

[0180] Furthermore, to ensure that the displayed cards can be accurately and without duplication, and can be successfully rendered on the homepage of the map application even in a weak network environment, this step uses an idempotent and reliable message push protocol for data transmission and rendering command distribution.

[0181] Message encapsulation and sending: The backend service encapsulates the generated display card data into a structured message containing a unique message ID and version number, and pushes it to the client through a message queue.

[0182] Idempotency check: After receiving a message, the client first checks the locally cached message processing records. If it finds that the unique message ID has already been processed (to prevent duplicate pushes due to network jitter), the message is discarded to avoid duplicate cards appearing on the homepage; if it has not been processed, the rendering logic continues to execute.

[0183] Reliability Guarantee: After completing card rendering, the client sends an ACK (acknowledgment) signal to the server. If the server does not receive the ACK within a specified time (e.g., the user is in an elevator or other environment without network access), a retry mechanism will be triggered, and the message will be re-pushed according to an exponential backoff strategy until acknowledgment is received.

[0184] Homepage rendering: After successful verification, the client parses the message body, calls the rendering engine of the map application homepage, and inserts the display card into the specified feed stream position or overlay area to complete the final visualization display.

[0185] In this implementation, by automatically matching differentiated card templates based on content volume and materials, the recommended content and visual form are accurately adapted, improving the efficiency and aesthetics of information display. At the same time, combined with an idempotent and reliable message push protocol, the problem of duplicate card display or loss caused by network anomalies is effectively solved, ensuring the consistency and stability of the user experience on the map application homepage.

[0186] In some optional implementations of this embodiment, the execution entity may also perform the following operations: in response to the failure condition that the target recommended topic in the display card reaches the non-immediate travel intention corresponding to the target recommended topic, the recommended topic is hidden in the display card; in response to all recommended topics in the display card reaching their respective failure conditions, the display card is stopped from being displayed.

[0187] Once the display card is successfully rendered on the preset display terminal (such as the homepage of a map application), the system backend will create an independent monitoring instance for each recommended topic in the card. This instance records the binding relationship between the currently displayed recommended topic and its source non-real-time travel intent, as well as the context parameters when the non-real-time travel intent was generated (such as generation timestamp, associated geofence range, specific search keywords, etc.). The system starts a lightweight status monitoring service to continuously or periodically compare the current real-time environmental data with the failure conditions of the non-real-time travel intent.

[0188] Failure conditions refer to a set of preset logical rules used to determine that non-immediate travel intentions no longer have recommendation value. In this embodiment, the determination logic for failure conditions includes, but is not limited to, the following situations: Spatiotemporal deviation determination: Obtain the user's real-time geographical location and current time. If the user has left the target area associated with the non-immediate travel intention (e.g., the user has left Beijing for Shanghai, while the original intention was "weekend trip to Beijing"), or the current time has exceeded the time window planned for the intention (e.g., the intention was "travel during the May Day holiday", but the current time has passed the May Day holiday), then the failure condition is determined to be met.

[0189] Explicit negative feedback detection: Monitor user interactions on the displayed cards. If a user performs an explicit negative feedback action such as "not interested," "close," or "long press to delete" on a recommended topic, the system immediately marks the action as a strong failure signal.

[0190] Task completion determination: If the system detects that the user has completed a task that is highly related to the non-immediate travel intention through other paths (for example, the user has just completed a flight ticket payment to a certain place, or checked in at the point of interest), then the intention is determined to have reached a closed loop and the failure condition is met.

[0191] Once the monitoring service determines that the non-real-time travel intent corresponding to a certain target recommendation topic meets the invalidation conditions, the system will immediately trigger a partial update instruction, including the following operations: Content Removal: Removes the invalid recommended theme from the data source of the currently rendering display card.

[0192] Layout Reflow: The front-end component receives the updated data list and reflows the remaining recommended themes. For example, if the original cards used a fluid layout, the remaining themes will automatically fill the empty spaces; if a fixed grid layout was used, the corresponding positions will be left empty or collapsed to ensure that outdated content is no longer displayed on the user interface. This process is smooth for the user and does not cause a full page refresh.

[0193] After performing the partial blocking operation, the system will further check the number of recommended topics remaining in the currently displayed card, including the following operations: Null value detection: If the system detects that the set of recommended topics belonging to this display card is empty (i.e. all recommended topics have reached their respective failure conditions), or the number of remaining topics is lower than the preset minimum display threshold (e.g., less than 1 valid topic), the system will determine that the display card has lost its value.

[0194] Card Unloading: The system sends a destruction command to the preset display terminal, completely removing the displayed card from the home view of the map application (Stop Displaying). At this time, the screen space originally occupied will be automatically filled by the content flow below, or restored to the default empty placeholder, thus avoiding the display of blank or invalid UI containers.

[0195] In this implementation, by monitoring the invalidation conditions of non-immediate travel intentions in real time and dynamically blocking or stopping the display of corresponding cards, expired invalid information due to changes in time and space or changes in user behavior can be cleared in a timely manner, avoiding visual interference to users and significantly improving the freshness of information and user experience on the map application homepage.

[0196] Continue to refer to Figure 4 The illustration shows a schematic flow 400 of yet another embodiment of the large-model-based topic recommendation method according to this disclosure. Flow 400 includes the following steps: Step 401: For objects that have issued non-real-time travel search requests within the first time period ending now, combine the spatiotemporal context of the objects' historical search requests within the second time period ending now to obtain a context sequence.

[0197] The duration of the first time period is shorter than the duration of the second time period.

[0198] Step 402: Convert the spatiotemporal quantized data in the spatiotemporal context into spatiotemporal semantic data to obtain a converted data pair including spatiotemporal quantized data and spatiotemporal semantic data.

[0199] Step 403: Based on the context sequence, transformation data pairs, object profile data, and prior guidance data, determine the object's non-immediate travel intention and at least one of the following: the clarity, priority, and failure condition of the non-immediate travel intention.

[0200] Step 404: Combine each non-immediate travel intention with at least one of the following: the clarity, priority, and invalidation condition of the non-immediate travel intention, to obtain a set of non-immediate travel intentions.

[0201] Step 405: Using a large artificial intelligence model, determine multiple recommendation topics based on prior information about the interest points corresponding to each non-immediate travel intention in the non-immediate travel intention set.

[0202] Step 406: Based on at least one of the following factors, such as the matching degree between the decision-making stage of the recommended topic and the object, the fit between the recommended topic and the current spatiotemporal scene, and the coverage breadth of the recommended topic, sort multiple recommended topics to obtain a set of recommended topics.

[0203] Step 407: For recommended topics in the recommended topic set, determine the target card template to which the recommended topic belongs from the card template set based on the content volume information and image material information of the recommended topic.

[0204] Step 408: Generate display cards based on the target card template and at least one recommended topic belonging to the target card template.

[0205] Step 409: Display the display card on the homepage of the map application via a preset display terminal.

[0206] The process 400 of the topic recommendation method based on the large model in this embodiment, compared with the above process 200, specifically describes the process of determining the intent set, the process of determining the recommended topic, and the process of displaying the recommended topic. By using the artificial intelligence large model to recommend non-real-time travel topics in map applications, the accuracy of topic recommendation results is improved, and the browsing needs of the object in the non-real-time travel stage are met.

[0207] Continue to refer to Figure 5 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a topic recommendation device based on a large model. This system embodiment is similar to... Figure 2 Corresponding to the method embodiments shown, the system can be specifically applied to various electronic devices.

[0208] like Figure 5 As shown, the large-scale model-based topic recommendation device 500 includes: a sequence generation unit 501, configured to obtain a context sequence for an object that has issued a non-immediate travel search request, by combining the spatiotemporal context of the object's historical search requests; an intent determination unit 502, configured to determine the object's non-immediate travel intent set based on the context sequence using an artificial intelligence large-scale model; and a topic determination unit 503, configured to determine the set of recommended topics to be displayed in the object's map application based on the non-immediate travel intent set using an artificial intelligence large-scale model.

[0209] In some optional implementations of this embodiment, the intent determination unit 502 is further configured to: convert spatiotemporal quantization data in the spatiotemporal context into spatiotemporal semantic data to obtain a converted data pair including spatiotemporal quantization data and spatiotemporal semantic data; and determine the set of non-real-time travel intents of the object based on the context sequence and the converted data pair using an artificial intelligence big data model.

[0210] In some optional implementations of this embodiment, the intent determination unit 502 is further configured to: determine the set of non-real-time travel intentions of an object by means of an artificial intelligence big data model, based on context sequence, conversion data pairs, object profile data and prior guidance data.

[0211] In some optional implementations of this embodiment, the intent determination unit 502 is further configured to: determine the object's non-immediate travel intent and at least one of the clarity, priority, and failure conditions of the non-immediate travel intent based on the context sequence, conversion data pairs, profile data, and prior guidance data; and obtain a set of non-immediate travel intents by combining each non-immediate travel intent and at least one of the clarity, priority, and failure conditions of the non-immediate travel intent.

[0212] In some optional implementations of this embodiment, the intent determination unit 502 is further configured to determine the priority in the following way: by using an artificial intelligence big data model, the priority is determined based on at least one of the following: the timeliness and clarity of the non-immediate travel intent, the feedback data of the object to the historical recommended topics corresponding to the non-immediate travel intent, and the fit between the non-immediate travel intent and the current spatiotemporal scene.

[0213] In some optional implementations of this embodiment, the topic determination unit 503 is further configured to: determine a set of recommended topics based on prior information data of interest points corresponding to each non-immediate travel intention in the non-immediate travel intention set and the non-immediate travel intention set using an artificial intelligence big data model.

[0214] In some optional implementations of this embodiment, the topic determination unit 503 is further configured to: determine multiple recommended topics based on prior information data of interest points corresponding to each non-immediate travel intention in the non-immediate travel intention set using an artificial intelligence big data model; and sort the multiple recommended topics to obtain a set of recommended topics based on at least one of the following: the matching degree between the recommended topics and the decision-making stage of the object, the fit between the recommended topics and the current spatiotemporal scene, and the coverage breadth of the recommended topics.

[0215] In some optional implementations of this embodiment, the above-mentioned device further includes: a theme display unit (not shown in the figure), configured to: for a recommended theme in the recommended theme set, determine the target card template to which the recommended theme belongs from the card template set based on the content volume information and image material information of the recommended theme; generate a display card based on the target card template and at least one recommended theme belonging to the target card template; and display the display card on the homepage of the map application through a preset display terminal.

[0216] In some optional implementations of this embodiment, the topic display unit is further configured to: block the recommended topic in the display card in response to the failure condition that the target recommended topic in the display card reaches the non-immediate travel intention corresponding to the target recommended topic; and stop displaying the display card in response to all recommended topics in the display card reaching their respective failure conditions.

[0217] In some optional implementations of this embodiment, the sequence generation unit 501 is further configured to: for an object that issues a non-instant travel search request within a first time period ending up to the present, combine the spatiotemporal context of the object's historical search requests within a second time period ending up to the present to obtain a context sequence, wherein the duration of the first time period is less than the duration of the second time period.

[0218] In this embodiment, a topic recommendation device based on a large model is provided. The sequence generation unit obtains a context sequence for an object that issues a non-immediate travel search request by combining the spatiotemporal context of the object's historical search requests. The intent determination unit determines the object's non-immediate travel intent set based on the context sequence using an artificial intelligence large model. The topic determination unit determines the recommended topic set to be displayed in the object's map application based on the non-immediate travel intent set using an artificial intelligence large model. By recommending non-immediate travel topics in the map application with the help of the artificial intelligence large model, the accuracy of topic recommendation results is improved, and the browsing needs of the object during the non-immediate travel stage are met.

[0219] According to embodiments of this disclosure, this disclosure also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the large-model-based topic recommendation method described in any of the above embodiments to be implemented.

[0220] According to embodiments of this disclosure, this disclosure also provides a readable storage medium storing computer instructions that, when executed by a computer, enable the large-model-based topic recommendation method described in any of the above embodiments.

[0221] This disclosure provides a computer program product that, when executed by a processor, can implement the large-model-based topic recommendation method described in any of the above embodiments.

[0222] Figure 6 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0223] like Figure 6 As shown, device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or a computer program loaded into random access memory (RAM) 603 from storage unit 608. RAM 603 may also store various programs and data required for the operation of device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0224] Multiple components in device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0225] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as the large model-based topic recommendation method. For example, in some embodiments, the large model-based topic recommendation method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the large model-based topic recommendation method described above can be performed. Alternatively, in other embodiments, the computing unit 601 can be configured to perform the large model-based topic recommendation method by any other suitable means (e.g., by means of firmware).

[0226] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0227] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable model-based subject recommendation device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0228] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0229] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0230] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0231] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, also known as cloud computing servers or cloud hosts, which are hosting products within the cloud computing service system to address the management difficulties and weak business scalability inherent in traditional physical hosts and Virtual Private Servers (VPS) services; they can also be servers for distributed systems or servers incorporating blockchain technology.

[0232] According to the technical solution of this disclosure embodiment, a topic recommendation device based on a large model is provided. The sequence generation unit, for an object issuing a non-immediate travel search request, combines the spatiotemporal context of the object's historical search requests to obtain a context sequence. The intent determination unit, through an artificial intelligence large model, determines the object's non-immediate travel intent set based on the context sequence. The topic determination unit, through an artificial intelligence large model, determines the recommended topic set to be displayed in the object's map application based on the non-immediate travel intent set. By using the artificial intelligence large model to recommend non-immediate travel topics in the map application, the accuracy of topic recommendation results is improved, meeting the browsing needs of the object during the non-immediate travel phase.

[0233] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution provided in this disclosure can be achieved, and this is not limited herein.

[0234] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A topic recommendation method based on a large model, comprising: For an object that issues a non-real-time travel search request, a context sequence is obtained by combining the spatiotemporal context of the object's historical search requests; Using a large artificial intelligence model, the set of non-real-time travel intentions of the object is determined based on the context sequence; The AI ​​big data model determines the set of recommended topics to be displayed in the object's map application based on the set of non-real-time travel intentions.

2. The method according to claim 1, wherein, The process of determining the set of non-real-time travel intentions of the object based on the context sequence using a large artificial intelligence model includes: The spatiotemporal quantized data in the spatiotemporal context is converted into spatiotemporal semantic data to obtain a converted data pair including the spatiotemporal quantized data and the spatiotemporal semantic data. Using the aforementioned large-scale artificial intelligence model, the non-real-time travel intention set of the object is determined based on the context sequence and the transformation data pairs.

3. The method according to claim 2, wherein, The step of determining the object's non-real-time travel intention set through the artificial intelligence big data model, based on the context sequence and the conversion data pair, includes: Using the aforementioned large-scale artificial intelligence model, the non-real-time travel intention set of the object is determined based on the context sequence, the conversion data pairs, the object's profile data, and prior guidance data.

4. The method according to claim 3, wherein, The step of determining the non-real-time travel intent set of the object based on the context sequence, the conversion data pair, the object's profile data, and prior guidance data includes: Based on the context sequence, the conversion data pair, the profile data, and the prior guidance data, determine the object's non-immediate travel intention and at least one of the following: the clarity, priority, and failure condition of the non-immediate travel intention; The set of non-immediate travel intentions is obtained by combining each of the aforementioned non-immediate travel intentions with at least one of the following: clarity, priority, and failure condition of the non-immediate travel intention.

5. The method according to claim 4, wherein, The priority is determined in the following manner: The priority is determined by the AI ​​big data model based on at least one of the following: the timeliness and clarity of the non-immediate travel intention, the feedback data of the object to the historical recommended topics corresponding to the non-immediate travel intention, and the degree of fit between the non-immediate travel intention and the current spatiotemporal scene.

6. The method according to any one of claims 1-5, wherein, The process of determining the set of recommended topics to display in the map application for the object based on the set of non-real-time travel intentions using the large artificial intelligence model includes: The recommended topic set is determined by the artificial intelligence big data model based on the set of non-real-time travel intentions and the prior information data of the interest points corresponding to each non-real-time travel intention in the set of non-real-time travel intentions.

7. The method according to claim 6, wherein, The process involves using the AI ​​big data model to determine the recommended topic set based on the set of non-real-time travel intentions and prior information about the interest points corresponding to each non-real-time travel intention in the set of non-real-time travel intentions. This includes: Using the aforementioned large-scale artificial intelligence model, multiple recommended topics are determined based on the set of non-real-time travel intentions and the prior information data of the interest points corresponding to each non-real-time travel intention in the set of non-real-time travel intentions. Based on at least one of the following factors—the matching degree between the recommended topic and the decision-making stage of the object, the fit between the recommended topic and the current spatiotemporal scene, and the coverage breadth of the recommended topic—multiple recommended topics are sorted to obtain the recommended topic set.

8. The method according to any one of claims 1-7, wherein, Also includes: For each recommended topic in the recommended topic set, the target card template to which the recommended topic belongs is determined from the card template set based on the content volume information and image material information of the recommended topic. Based on the target card template and at least one of the recommended topics belonging to the target card template, a display card is generated; The display card is shown on the homepage of the map application via a preset display terminal.

9. The method according to claim 8, wherein, Also includes: In response to a failure condition that the target recommended topic in the display card meets the non-immediate travel intention corresponding to the target recommended topic, the recommended topic is hidden in the display card; When all recommended topics in the display card have met their respective failure conditions, the display card is stopped from being displayed.

10. The method according to claim 1, wherein, For an object that issues a non-real-time travel search request, a context sequence is obtained by combining the spatiotemporal context of the object's historical search requests, including: For an object that issued the non-instant travel search request within the first time period ending now, the context sequence is obtained by combining the spatiotemporal context of the object's historical search requests within the second time period ending now, wherein the duration of the first time period is shorter than the duration of the second time period.

11. A topic recommendation device based on a large model, comprising: The sequence generation unit is configured to, for an object that issues a non-real-time travel search request, combine the spatiotemporal context of the object's historical search requests to obtain a context sequence; The intent determination unit is configured to determine the set of non-real-time travel intentions of the object based on the context sequence using a large artificial intelligence model. The theme determination unit is configured to determine, through the artificial intelligence big data model, a set of recommended themes to be displayed in the object's map application based on the set of non-real-time travel intentions.

12. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-10.

13. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-10.

14. A computer program product comprising: A computer program that, when executed by a processor, implements the method according to any one of claims 1-10.

Images