Multimodal large model active recommendation system based on LBS

Abstract

The present invention discloses a multimodal large model active recommendation system based on LBS, which specifically relates to the field of tourism recommendation systems and is used to solve the problem of mismatched recommendations caused by multimodal data dislocation. It is achieved by comprehensively collecting user historical behavior data and identifying key geographical locations, and dynamically adjusting the center point, influence radius and weight parameters of the field of interest in combination with advanced algorithms to ensure that the recommended content accurately fits user needs and adapts to behavioral changes. Through the LBS field of interest mechanism, the problem of "multimodal data dislocation effect" in urban peripheral tourism recommendations is effectively solved, avoiding mismatched recommendations caused by long-term and short-term interest mismatches in traditional systems. Real-time monitoring and adjustment of the scope and weight of the field of interest ensure that the recommended content is highly matched with the user's location and needs. At the same time, the trigger value constructed by the degree of location anomaly and the frequency of interest switching is used to accurately identify changes in user behavior and timely adjust the recommendation strategy, thereby improving the accuracy, personalization and user satisfaction of recommendations.

Description

Technical Field

[0001] The present invention relates to the field of tourism recommendation systems, and more specifically, to a multi-modal large model active recommendation system based on LBS. Background Art

[0002] In the scenario of city peri-city travel recommendations, users' interests and needs often change significantly with changes in their geographic location and adjustments in their immediate behavior. For example, when users move from the city center into the suburbs or other tourist attractions, their interests may shift from shopping and dining to natural landscapes and outdoor activities. In this scenario, the recommendation system needs to make accurate content recommendations based on the user's location and real-time needs to ensure that users can get travel recommendations that match their current interests.

[0003] However, existing recommendation systems often face the problem of "multimodal data misalignment" when dealing with such dynamic changes. Traditional systems usually rely on fixed user interest models and cannot promptly reflect new interests of users due to changes in location and immediate needs. This causes the recommended content to be out of touch with the actual needs of users, especially when users frequently switch activity areas. For example, when a user moves from a shopping area to a natural scenic area, the traditional system may still recommend shopping-related content, but cannot quickly adjust the recommendation strategy to provide outdoor activity suggestions that are more in line with the user's current situation. This mismatch not only reduces the user experience, but may also cause users to distrust the recommendation system. Existing systems have significant deficiencies in real-time adjustment and dynamic adaptation of interest scenarios, making it difficult to meet users' personalized recommendation needs in different geographical locations and immediate needs.

[0004] In order to solve the above problems, a technical solution is now provided. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiment of the present invention provides a multimodal large model active recommendation system based on LBS, which achieves a significant improvement in the accuracy and personalization level of the recommendation system by comprehensively collecting user historical behavior data, identifying key geographical locations, and combining advanced algorithms to build and optimize interest fields. The dynamic adjustment of the center point, influence radius and weight parameters of the interest field ensures that the recommended content is more in line with the actual needs of users in terms of geographical distribution and time, and enhances the system's adaptability to changes in user behavior. Through the LBS interest field mechanism, the system can effectively solve the problem of multimodal data dislocation effect in urban peripheral tourism recommendations, and avoid the mismatch of long-term and short-term interests in traditional recommendation systems, resulting in recommended content that does not meet user needs. On the basis of real-time acquisition of user location and synchronous update of interest field status, the system can dynamically match the user's actual location to ensure that the recommended content is accurate and timely. Through the fusion analysis of short-term interests and interest fields, the system can improve the matching degree of recommended content with users' current needs, and ensure the diversity and comprehensiveness of recommended content by balancing the interaction between multiple interest fields. In addition, by using the trigger values ​​constructed by the degree of location anomaly and the frequency of interest switching, the system can accurately identify significant changes in user behavior and timely adjust the scope of interest and the priority of recommended content, thereby improving the accuracy and personalization of recommendations, and ultimately improving user experience and satisfaction. This solves the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The LBS-based multi-modal large-model active recommendation system includes: interest field creation module, location synchronization module, interest fusion module and dynamic adjustment module;

[0008] Interest field creation module: Create multiple interest fields, define the center point, influence radius and weight of the interest field based on the user's long-term behavior data, cover the main activity areas around the city, and output the interest field information to the location synchronization module.

[0009] Position synchronization module: synchronizes user position and interest field status in real time, adjusts the influence range and priority of the interest field to ensure consistency with the user's movement path, and outputs the synchronized interest field status to the interest fusion module.

[0010] Interest fusion module: integrates and analyzes short-term interests with current interest fields, dynamically adjusts interest field weights, balances the interactions among multiple interest fields, and outputs the fused interest field weights to the dynamic adjustment module.

[0011] Dynamic adjustment module: It evaluates behavioral changes by building a trigger value based on the abnormality of the user's geographic location and the frequency of switching interest areas, dynamically adjusts the scope of influence of the interest field, recalculates the influence radius and center point, and updates the priority of recommended content.

[0012] In a preferred embodiment, the operation process of the interest field creation module is as follows:

[0013] S1.1, comprehensively collect users’ historical travel records; classify users’ preference types and summarize them into major areas of interest.

[0014] S1.2, in the urban periphery, LBS technology is used to identify key geographical locations that users have visited many times in the past.

[0015] S1.3, construct corresponding fields of interest based on the user's preference type and the identified key geographical locations; each field of interest is composed of several geographical location points, forming a basic structure of a field of interest in geographic space; to ensure that the field of interest can accurately cover the user's activity area, define the center point, influence radius and related weight parameters for each field of interest during initialization.

[0016] S1.4. In the process of constructing the field of interest, multivariate regression analysis is used to calculate the comprehensive weight of user behavior data, and the field of interest is spatially optimized through weighted Voronoi diagram partitioning technology, and environmental data is introduced into the graph structure model; according to the user's preference type and behavior pattern at each location, the influence radius of each field of interest is dynamically adjusted, and the gravity model is used to calculate the interaction parameters between different fields of interest, so as to optimize the spatial layout of the field of interest.

[0017] S1.5, after the creation and initialization of the field of interest, use the path planning algorithm to simulate the path analysis of the user's recent behavior data; combine the geographical weighted regression model to correct the center point and influence radius of the field of interest to ensure that the field of interest is highly consistent with the user's long-term behavior pattern; during the correction process, predict the trend of user behavior changes through time series analysis, and use the Kalman filter algorithm to dynamically adjust the weight parameters.

[0018] In a preferred embodiment, the operation process of the position synchronization module is as follows:

[0019] S2.1, when the user enters the tourist area around the city, the user's current location is obtained in real time through LBS technology; the location data is expressed in the form of geographic coordinates; to ensure the accuracy of the data, the Kalman filter is used to filter the noise and smooth the trajectory of the location data to obtain the user's location.

[0020] S2.2, after obtaining the user's current location, calculate the geographic distance between the user's location and the center point of each field of interest; use the Haversine formula in the earth coordinate system to calculate the distance to reflect the impact of the earth's curvature on the distance.

[0021] S2.3, dynamically adjust the influence radius of the field of interest according to the distance between the user's current position and the center point of the field of interest and the user's moving speed in the corresponding area.

[0022] S2.4, based on the overlap between the user's position and the field of interest, the weight parameters of the field of interest are updated in real time, using the gravity model for calculation.

[0023] S2.5, after completing the adjustment of the influence radius and weight parameters of the field of interest, apply the updated parameters to the state of the field of interest, and synchronize the coverage of the field of interest in real time; during the user's movement, continue to repeat the calculation steps of steps S2.1-S2.5 to ensure that the spatial layout and weight parameters of the field of interest are consistent with the user's actual position and behavior dynamics.

[0024] In a preferred embodiment, the operation process of the interest fusion module is as follows:

[0025] S3.1, real-time acquisition of user behavior data, including recent social interactions and instant search behaviors; recorded in the form of timestamps, social interaction data is , instant search data is ,in Indicates the time point of obtaining data; use time series analysis technology to smooth the behavior data, remove outliers, and obtain the trend function of the user's current short-term interest and .

[0026] S3.2 Converting users’ real-time behavior data into short-term interest influence coefficients , reflects the influence of short-term interests on the current interest field; the specific calculation formula is as follows: ;in, is the time decay coefficient, which is used to control the timeliness of short-term interests; Indicates a point in time in the past; and They are the trend functions of social interaction and instant search respectively.

[0027] S3.3, based on the calculated short-term interest influence coefficient, perform a fusion analysis with the user's current interest field; let the weight of the user's current interest field be , then the weight of the fused interest field is Dynamic adjustment is performed using the following formula: ;in, It is the sensitivity parameter of the interest field, reflecting the responsiveness of the interest field to short-term interests.

[0028] S3.4, while adjusting the weight of the current field of interest, balance the interactions between multiple fields of interest; use the interaction balance formula based on the gravity model: ;in, For interest field Field of interest The interaction forces between is the distance between the center points of the two fields of interest; and are the short-term interest influence coefficients of the current interest field and other interest fields, respectively.

[0029] S3.5, applies the interest field weights after fusion analysis to the recommendation algorithm, updates the recommended content in real time, and pushes the travel destinations most relevant to the user’s current interest field.

[0030] In a preferred embodiment, the operation process of the dynamic adjustment module is as follows:

[0031] Position anomaly index generation logic:

[0032] A1, collects the user's location data within a specific time period and records it as a position vector sequence ,in and Represents the user at a certain point in time Longitude and latitude; standardize the location data and use normalization transformation to obtain the standardized position vector .

[0033] A2, based on the standardized position vector, construct the covariance matrix of the position data , which is used to describe the distribution characteristics of position data; the calculation formula of the covariance matrix is: ;in, is the total number of location data points; is the mean vector of the normalized position vector; Represents the position vector after removing the mean.

[0034] A3: Use Mahalanobis distance to measure the abnormality of the standardized position vector and calculate the abnormality score of each position point. , used to identify abnormal location behavior: ;in, is the mean vector of the normalized position vector; is the inverse matrix of the covariance matrix; Representing time point The location anomaly score.

[0035] A4, based on the anomaly score at each time point, calculates the location anomaly index to comprehensively evaluate the degree of anomaly of user location data over a period of time; this is accomplished by accumulating anomaly scores and introducing a time decay factor , to reflect the timeliness of location anomalies: ;in, is the time decay factor, which controls the decay speed of the anomaly score; is the total time period for calculation; A point in time in the past.

[0036] In a preferred embodiment, the interest transition frequency generation logic is:

[0037] B1. First, collect user behavior data within a specific time period. ,in Indicates a time point; the behavior data within a specific time period is classified according to different areas of interest, and the behavior data in each area is represented as a vector ,in Indicates Areas of interest.

[0038] B2, conduct time series analysis on the behavior data of each interest area to identify the changing trends of users' interests in different time periods; use discrete wavelet transform to decompose the time series data to obtain different frequency components ,in Indicates the level of the frequency component; ;in, is the wavelet basis function; for time point; is the total number of time points.

[0039] B3, within each area of ​​interest, by calculating the interest transition rate at adjacent time points , identify the interest transition point, and the interest transition rate is: ;in, Indicate area of ​​interest At the point in time The transition rate of is the frequency component after wavelet decomposition.

[0040] B4, for all interest areas, accumulate the frequency of interest transitions within a period of time to generate the interest transition frequency ; In the accumulation process, the time weighting factor is introduced , in order to control the timeliness of the transition; the calculation formula is: ;in, is the total number of areas of interest; is the observation period; for a point in time in the past; is the time weighting factor, which controls the timeliness of the transition.

[0041] In a preferred embodiment, the location anomaly index and interest transition frequency are comprehensively processed using a weighted harmonic average method to obtain a trigger value; if the trigger value is greater than or equal to the trigger threshold, it means that the user's current geographic location anomaly and interest change frequency have reached a critical level requiring a system response; this means that the user's behavior pattern has changed significantly. In this case, it is necessary to immediately and dynamically adjust the range of the field of interest, recalculate the influence radius and the center position, to ensure that the field of interest can accurately cover the user's current activity area; at the same time, update the priority of the recommended content.

[0042] In a preferred embodiment, when updating the priority of the recommended content, the weight of each recommended content is adjusted by comprehensively analyzing the latest distance between the user's current location and the center point of the field of interest, behavioral feedback, and the size of the trigger value.

[0043] The technical effects and advantages of the LBS-based multi-modal large model active recommendation system of the present invention are as follows:

[0044] 1. Through the comprehensive collection of user historical behavior data, the identification of key geographical locations, and the use of advanced algorithms to build and optimize interest fields, the accuracy and personalization of the recommendation system can be significantly improved. The dynamic adjustment of the center point, influence radius and weight parameters of the interest field makes the recommended content more in line with the actual needs of users in terms of geographical distribution and time. Furthermore, by simulating path analysis and correcting the layout of the interest field, the system is ensured to be highly adaptable to future changes in user behavior, thereby optimizing the recommendation effect and improving user satisfaction.

[0045] 2. By acquiring the user's location in real time and updating the status of the field of interest synchronously, it can ensure that the influence range of the field of interest dynamically matches the user's actual location, avoiding inaccurate recommendations due to changes in geographic location. Dynamically adjusting the influence radius and weight parameters of the field of interest not only improves the accuracy and personalization of recommended content, but also enhances the system's ability to adapt to changes in user behavior. Ultimately, this process ensures that the recommendation system can provide high-quality content that meets the user's current interests and needs at every moment of the user's movement.

[0046] 3. By integrating short-term interests with interest fields and dynamically adjusting the weights of interest fields, the matching degree between recommended content and users' current needs can be significantly improved. Balancing the interaction between multiple interest fields and avoiding excessive dominance of recommendations by a single interest field can make the recommended content more diverse and comprehensive. Ultimately, this process enhances the recommendation system's ability to respond to changes in user behavior, ensuring that recommended content can always remain highly consistent with user interests, thereby improving user satisfaction and experience.

[0047] 4. By using the trigger values ​​constructed by the degree of location anomaly and the frequency of interest switching, the system can accurately identify significant changes in user behavior, thereby timely adjusting the scope of interest fields and the priority of recommended content. This ensures that the system can quickly provide the most relevant content when user needs change, improve the accuracy and personalization of recommendations, and ultimately improve user experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a structural diagram of the multi-modal large model active recommendation system based on LBS of the present invention;

[0049] Figure 2 It is a flow chart of the interest field creation module of the LBS-based multimodal large model active recommendation system of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Example 1

[0052] Figure 1 The invention provides a multi-modal large model active recommendation system based on LBS, comprising: an interest field creation module, a location synchronization module, an interest fusion module and a dynamic adjustment module.

[0053] Interest field creation module: Create multiple interest fields, define the center point, influence radius and weight of the interest field based on the user's long-term behavior data, cover the main activity areas around the city, and output the interest field information (center point, influence radius, weight) to the location synchronization module.

[0054] Location synchronization module: synchronizes user location and interest field status in real time, adjusts the influence range and priority of the interest field to ensure consistency with the user's movement path, and outputs the synchronized interest field status (real-time location, adjusted influence range, priority) to the interest fusion module.

[0055] Interest fusion module: integrates and analyzes short-term interests with current interest fields, dynamically adjusts interest field weights, balances the interactions among multiple interest fields, and outputs the fused interest field weights to the dynamic adjustment module.

[0056] Dynamic adjustment module: It evaluates behavioral changes by building a trigger value based on the abnormality of the user's geographic location and the frequency of switching interest areas, dynamically adjusts the scope of influence of the interest field, recalculates the influence radius and center point, and updates the priority of recommended content.

[0057] Long-term interest refers to the stable interest tendency shown by users over a long period of time. This interest is usually a deep-level interest field extracted from the user's continuous behavior patterns, historical access records, preference habits and other data. This interest has high persistence and regularity, reflecting the user's long-term preference and habitual choice in a specific field.

[0058] Short-term interests refer to the interest tendencies that users show in a specific period of time due to temporary needs, environmental changes, or external stimuli. This interest is often highly immediate and volatile, and is easily affected by factors such as the user's current situation, social interaction, and instant search. Short-term interests reflect users' immediate needs and short-term behavioral trends at a specific moment.

[0059] In order to provide highly personalized recommendations for travel recommendations around cities, it is necessary to conduct an in-depth analysis of the user's historical travel behavior and combine it with geographic location data to build an interest field that can reflect the user's long-term interests. Through the detailed classification of the user's past behavior and the accurate identification of key geographic locations, the interest field not only covers the user's main activity area, but also dynamically adapts to the user's immediate needs. This process lays the foundation for the recommendation system to be highly consistent with the user's behavior pattern in geographic space, ensuring that the recommended content can more accurately meet the user's interests and preferences.

[0060] like Figure 2 As shown in the figure, the operation process of the interest field creation module is as follows:

[0061] S1.1, comprehensively collect users’ historical travel records, including the attractions they have visited, length of stay, frequency of visits, etc. Classify users’ preference types and group them into major areas of interest, such as culture (museums, historical sites), nature (parks, nature reserves), shopping (shopping malls, pedestrian streets), etc. During the classification process, refine the users’ behavior patterns and their interactive behaviors in various attractions (such as evaluation, sharing, and interaction on social media).

[0062] S1.2, in the surrounding areas of the city, use LBS technology to identify key geographical locations that users have visited many times in the past. These locations include attractions that users frequently visit, places of special significance (such as famous tourist check-in spots), and potential attractions that match the user's preference types. Based on the distribution of these locations, they are mapped to the corresponding areas of interest to ensure the match between the geographical locations and the types of interests. When identifying key geographical locations, it can be based on the frequency of users' visits to a certain place within a certain period of time. When the number of visits to a certain geographical location by a user in the past six months exceeds the average of all the locations visited by the user by more than 1.5 times, the location can be regarded as a key geographical location. The setting of this threshold is based on the average visit level within the rated time, and statistical analysis of the user's historical behavior data is used to distinguish those places with higher attention to ensure the representativeness and importance of these locations in the field of interest.

[0063] S1.3, construct corresponding fields of interest based on the user's preference type and the identified key geographical locations. Each field of interest consists of several geographical location points, which form the basic structure of a field of interest in geographic space. To ensure that the field of interest can accurately cover the user's possible activity area, the center point, influence radius and related weight parameters are defined for each field of interest during initialization. The center point is set based on the location that the user visits most frequently, the influence radius takes into account the user's range of movement in this type of activity and the correlation between places, and the weight parameters are initially allocated based on the user's preference strength.

[0064] Determination of the center point: Use a density-based spatial clustering algorithm (such as DBSCAN) to perform cluster analysis on the user's historical access data to identify the most frequently visited locations in a specific place of interest. By analyzing the degree of clustering of these locations, the area with the highest density is extracted as the center point. If the number of visits in a cluster exceeds a preset threshold (for example, the number of visits to the location by the user in the past year exceeds 20% of the total number of visits), then the location is set as the center point of the field of interest.

[0065] Calculation of influence radius: Based on the actual movement path of users in this type of activity, the buffer analysis technology (BufferAnalysis) in the geographic information system (GIS) is used to calculate the average distance that users move outward from the center point. The specific value of the influence radius is determined by the average movement range of users in this type of activity, and the correlation between different geographical locations is taken into account. The setting of the buffer radius combines the maximum and minimum movement distances of users, and takes 80% of the movement range of users in this type of activity as the radius to ensure coverage of the main activity area.

[0066] Initial allocation of weight parameters: The weighted average method is used to initially allocate weight parameters based on the user's preference strength for different places of interest. The preference strength is obtained by analyzing the user's stay time, visit frequency, and social interaction data at each location. Using regression analysis technology, the user's behavior data at a specific location is mapped to a numerical weight, and the priority of different places of interest is predicted through a linear regression model. Finally, these numerical weights are normalized and used as the initial weight parameters of the interest field in the recommendation strategy.

[0067] S1.4, in the process of constructing the field of interest, multivariate regression analysis is used to calculate the comprehensive weight of user behavior data (visit frequency, length of stay, social interaction), and the field of interest is spatially optimized through weighted Voronoi diagram partitioning technology, and environmental data (such as traffic convenience and distance between attractions) are introduced into the graph structure model. According to the user's preference type and behavior pattern at each location, the influence radius of each field of interest is dynamically adjusted, and the interaction parameters between different fields of interest are calculated using the gravity model, so as to optimize the spatial layout of the field of interest so that it can accurately reflect the user's actual interest area and its association.

[0068] S1.5, after the creation and initialization of the field of interest, use the path planning algorithm (such as Dijkstra algorithm) to simulate the path analysis of the user's recent behavior data to evaluate the coverage of the field of interest on the user's behavior. Combined with the geographically weighted regression (GWR) model, calibrate the center point and influence radius of the field of interest to ensure that the field of interest is highly consistent with the user's long-term behavior pattern. During the calibration process, predict the trend of user behavior changes through time series analysis, and use the Kalman filter algorithm to dynamically adjust the weight parameters to improve the accuracy of the field of interest and its adaptability to future behavior.

[0069] By comprehensively collecting historical user behavior data, identifying key geographic locations, and using advanced algorithms to build and optimize interest fields, the accuracy and personalization of the recommendation system can be significantly improved. The dynamic adjustment of the center point, influence radius, and weight parameters of the interest field makes the recommended content more in line with the actual needs of users in terms of geographical distribution and time. Furthermore, by simulating path analysis and correcting the layout of the interest field, the system is ensured to be highly adaptable to future changes in user behavior, thereby optimizing the recommendation effect and improving user satisfaction.

[0070] In the scenario of city peripheral tourism recommendation, the user's geographical location is constantly changing, and their behavior and needs are also dynamically adjusted during the movement. In order to ensure that the recommended content can meet the user's actual needs in real time, it is necessary to obtain the user's current location through LBS technology and synchronously update the user's field of interest status. By accurately calculating the distance between the user's location and the center point of the field of interest, and dynamically adjusting the influence radius and weight parameters of the field of interest, it can be ensured that the system can continue to provide the most relevant and appropriate recommended content when the user moves to different locations.

[0071] The operation process of the position synchronization module is as follows:

[0072] S2.1, when the user enters the tourist area around the city, the user's current location is obtained in real time through LBS technology. The location data is expressed in the form of geographic coordinates (longitude Lon, latitude Lat). To ensure the accuracy of the data, the Kalman filter is used to filter the noise and smooth the trajectory of the location data to obtain the user's location ,in Indicates a point in time.

[0073] :User at time point The geographical location consists of longitude and latitude.

[0074] S2.2, after obtaining the user's current location, calculate the user's location With the center point of each field of interest The geographical distance between two points. The Haversine formula in the earth coordinate system is used to calculate the distance to reflect the effect of the earth's curvature on the distance. The specific calculation formula is: ;

[0075] : Field of interest The center point, longitude and latitude.

[0076] : The distance between the user and the center point of the field of interest, calculated using the Haversine formula.

[0077] : The radius of the Earth, about 6371 kilometers.

[0078] This formula at each time point The distances between the user and all the fields of interest are calculated to determine which field of interest the user is currently in.

[0079] S2.3, based on the distance between the user's current position and the center point of the field of interest and the user's moving speed in the corresponding area (in is the time interval), dynamically adjust the influence radius of the field of interest The adjustment rules are: ;in, is the influence radius of the field of interest at the current time point, is an adjustment factor used to control the sensitivity of radius adjustment. This formula makes the influence radius gradually shrink when the user approaches the center of the field of interest, and vice versa, when the user moves away from the center, the influence radius expands moderately.

[0080] S2.4, based on the overlap between the user's location and the field of interest, update the weight parameters of the field of interest in real time , calculated using the gravity model. The core of weight update is to consider the mutual influence between the user's current interest field and other interest fields, and use the gravity model for calculation: ;in, It is the basic weight of the interest field, which is determined by the user's long-term behavior data; For interest field Other fields of interest The distance between the two is also calculated by the Haversine formula. The formula considers the interaction between multiple interest fields through the gravity model and dynamically adjusts the weight of the current interest field to ensure that the priority of the recommended content is consistent with the actual needs of the user.

[0081] S2.5, after completing the adjustment of the influence radius and weight parameters of the field of interest, apply the updated parameters to the state of the field of interest, and synchronize the coverage of the field of interest in real time. During the user's movement, the calculation steps of steps S2.1-S2.5 are continuously repeated to ensure that the spatial layout and weight parameters of the field of interest are consistent with the user's actual location and behavior dynamics, ensuring the high matching degree and real-time nature of the recommended content.

[0082] By obtaining the user's location in real time and updating the status of the field of interest synchronously, it is possible to ensure that the influence range of the field of interest dynamically matches the user's actual location, avoiding inaccurate recommendations due to changes in geographic location. Dynamically adjusting the influence radius and weight parameters of the field of interest not only improves the accuracy and personalization of recommended content, but also enhances the system's ability to adapt to changes in user behavior. Ultimately, this process ensures that the recommendation system can provide high-quality content that meets the user's current interests and needs at every moment of the user's movement.

[0083] In the scenario of city peripheral tourism recommendation, user interests will change with real-time behaviors (such as social interaction and instant search). In order to ensure that the recommended content can closely match the user's current needs, it is necessary to integrate and analyze the user's short-term interests with the current interest field. By capturing the user's real-time behavior data, dynamically adjusting the weight of the interest field, and balancing the interaction between multiple interest fields, it can accurately reflect the user's immediate interest changes and provide more flexible and personalized content support for the recommendation system.

[0084] The operation process of the interest fusion module is as follows:

[0085] S3.1, real-time acquisition of user behavior data, including recent social interactions (such as likes, comments, shares) and instant search behavior. These data are recorded in the form of timestamps, for example, social interaction data is , instant search data is ,in Indicates the time point of obtaining data. Use time series analysis technology to smooth the behavior data, remove outliers, and obtain the trend function of the user's current short-term interest. and .

[0086] S3.2 Converting users’ real-time behavior data into short-term interest influence coefficients , this coefficient reflects the influence of short-term interests on the current interest field. The specific calculation formula is as follows: ;

[0087] in, is the time decay coefficient, which is used to control the timeliness of short-term interests; Indicates a point in time in the past; and are the trend functions of social interaction and instant search, respectively. This formula calculates the influence coefficient of the user's current short-term interest on various activities (such as nature, culture, and shopping) by weighted integration of behavioral data in the past.

[0088] S3.3, based on the calculated short-term interest impact coefficient, perform a fusion analysis with the user's current interest field. Suppose the weight of the user's current interest field is , then the weight of the fused field of interest is Dynamic adjustment is performed using the following formula: ;in, is the sensitivity parameter of the interest field, reflecting the responsiveness of the interest field to short-term interests. Through this formula, when the user's short-term interest shows a high tendency towards a certain type of activity (such as outdoor activities), the weight of the relevant interest field will be increased to ensure that the recommended content is more in line with the user's current needs.

[0089] S3.4, while adjusting the weight of the current interest field, it is necessary to balance the interactions between multiple interest fields to prevent a single interest field from overly dominating the recommended content. The interaction balance formula based on the gravity model is used: ;in, For interest field Field of interest The interaction force between is the distance between the center points of the two fields of interest, and are the short-term interest influence coefficients of the current interest field and other interest fields, respectively. This formula ensures that when the influence of short-term interest is large, the weight of the relevant interest field will increase moderately, while the weights of other interest fields will be dynamically adjusted according to the interaction force.

[0090] S3.5, apply the interest field weight after fusion analysis to the recommendation algorithm, update the recommended content in real time, and push the travel places or activities most relevant to the user's current interest field. At the same time, monitor the user's feedback on the recommended content (such as clicks, stay time), and further adjust the interest field weight according to the feedback data to ensure that the recommended content continues to be consistent with user needs.

[0091] By integrating short-term interests with interest fields and dynamically adjusting the weights of interest fields, the matching degree between recommended content and users' current needs can be significantly improved. Balancing the interaction between multiple interest fields and avoiding excessive dominance of recommendations by a single interest field can make the recommended content more diverse and comprehensive. Ultimately, this process enhances the recommendation system's ability to respond to changes in user behavior, ensuring that recommended content can always maintain a high degree of consistency with user interests, thereby improving user satisfaction and experience.

[0092] In the scenario of city peripheral tourism recommendation, the user's geographical location and interests often change dynamically. These changes may trigger the user's demand for different fields of interest, so it is necessary to adjust the scope of the field of interest and the priority of content recommendation in real time. By comprehensively evaluating the abnormality of the user's geographical location and the switching frequency of the field of interest, the system can accurately capture the user's behavioral changes. When these changes reach a certain level, the system will trigger a dynamic adjustment mechanism to ensure that the recommended content is highly matched with the user's current needs.

[0093] The operation process of the dynamic adjustment module is as follows:

[0094] In the scenario of city peripheral tourism recommendation, the user's geographical location usually follows a certain behavior pattern, but in some cases, the user may deviate from his or her regular activity area and show abnormal geographical behavior. These abnormal behaviors may indicate that the user's interests or needs have changed significantly, so it is necessary to identify and respond to these abnormal phenomena through the location anomaly index (ALI). Through in-depth analysis of user location data, the location anomaly index can help the system accurately capture the user's abnormal behavior in geographical location, providing a basis for subsequent dynamic adjustments and personalized recommendations.

[0095] Position anomaly index generation logic:

[0096] A1, collects the user's location data within a specific time period and records it as a position vector sequence ,in and Represents the user at a certain point in time In order to eliminate the dimension difference of the position data, the position data is standardized and the normalized transformation is used to obtain the standardized position vector .

[0097] A2, based on the standardized position vector, construct the covariance matrix of the position data , which is used to describe the distribution characteristics of position data. The calculation formula of the covariance matrix is: ;in, is the total number of location data points; is the mean vector of the normalized position vector (to avoid complexity, this part can be replaced by a simple form); Represents the position vector after removing the mean.

[0098] A3: Use Mahalanobis distance to measure the abnormality of the standardized position vector and calculate the abnormality score of each position point. , used to identify abnormal location behavior: ;in, is the mean vector of the normalized position vector; is the inverse of the covariance matrix, Representing time point The location anomaly score.

[0099] A4, based on the anomaly score at each time point, calculates the location anomaly index (ALI) to comprehensively evaluate the degree of anomaly of user location data over a period of time. The calculation of ALI is completed by accumulating anomaly scores and introducing a time decay factor , to reflect the timeliness of location anomalies: ;in, is the time decay factor, which controls the decay speed of the anomaly score; is the total time period for calculation; A point in time in the past.

[0100] The location anomaly index can timely identify users' deviations from their geographic location through comprehensive analysis and anomaly detection of user location data. This recognition capability enables the recommendation system to respond quickly when user behavior is abnormal, adjust the coverage and priority of interest fields, and thus improve the adaptability and accuracy of recommended content. Ultimately, this process helps to provide more personalized recommendation services, ensuring that the system can provide content that matches users' actual interests when their needs change, thereby enhancing the overall user experience.

[0101] In the scenario of city peripheral tourism recommendation, users' interests are often not limited to a single field, but they will quickly switch between different interest fields, such as from cultural attractions to natural landscapes or shopping places. In order to capture this variability of users in a short period of time, it is necessary to identify and quantify the user's conversion between different interest fields through interest transition frequency. Through time series analysis and frequency calculation of user behavior data, interest transition frequency can reveal the user's current diversified needs and provide an accurate dynamic interest portrait for the recommendation system.

[0102] Interest transition frequency generation logic:

[0103] B1. First, collect user behavior data within a specific time period. ,in Indicates a time point. Behavioral data includes information such as user clicks, searches, and social interactions. Behavioral data within a specific time period is classified according to different interest areas (such as culture, nature, and shopping), and the behavioral data in each area is represented as a vector ,in Indicates Areas of interest.

[0104] B2, conduct time series analysis on the behavior data of each interest area to identify the changing trends of users' interests in different time periods. Use Discrete Wavelet Transform (DWT) to decompose the time series data to obtain different frequency components. ,in Indicates the level of the frequency component.

[0105] ;in, is the wavelet basis function; for time point; is the total number of time points. Through wavelet decomposition, the frequency components of the behavior data can be extracted to capture the multi-scale changes in user interests.

[0106] B3, within each area of ​​interest, by calculating the interest transition rate at adjacent time points , identify the interest transition point. Define the interest transition rate as: ;in, Indicate area of ​​interest At the point in time The transition rate of is the frequency component after wavelet decomposition. This formula obtains the rate of change of interest by taking the derivative and identifies the transition point of interest.

[0107] B4, for all interest areas, accumulate the frequency of interest transitions within a period of time to generate the interest transition frequency In the accumulation process, the time weighting factor is introduced , in order to control the timeliness of the transition. The calculation formula is: ;in, is the total number of areas of interest; is the observation period; for a point in time in the past; is the time weighting factor, which controls the timeliness of the transition. This formula accumulates the transition rates of various interest areas to obtain a comprehensive transition frequency, which reflects the variability of users' interests in a specific time period.

[0108] The frequency of interest transitions can timely identify rapid changes in user interests by analyzing the rate at which users switch between different interest areas. This recognition capability enables the recommendation system to quickly adjust the recommendation strategy when users' interests frequently transition, ensuring that the recommended content can reflect the user's actual interests in real time. By applying this indicator, the system can more accurately respond to the diverse needs of users, improve the personalization and adaptability of recommended content, and thus improve user satisfaction and overall experience.

[0109] The position anomaly index and the interest transition frequency are processed comprehensively using a weighted harmonic average method to obtain a trigger value. For example, the trigger value can be calculated using the following formula: ;in and are respectively the preset proportional coefficients of the position anomaly index and the interest transition frequency, and both are greater than zero; is the trigger value.

[0110] The trigger value calculated after comprehensive processing indicates the comprehensive degree of the user's geographical location abnormality and interest change frequency in the current scenario. The size of the trigger value directly reflects the dynamic changes in user behavior and whether the system needs to adjust the field of interest. When the trigger value is large, it means that the user's geographical location and interest field have changed significantly. The system needs to respond quickly and adjust the scope, center position and priority of the field of interest to better meet the user's current needs; when the trigger value is small, it means that the user's behavior is relatively stable, and the system can maintain the current recommendation strategy without major adjustments. This trigger value comprehensively considers the location anomaly index and interest transition frequency to ensure that the recommendation system can sensitively capture the user's abnormal behavior and interest changes, thereby providing more accurate and personalized content recommendations.

[0111] If the trigger value is greater than or equal to the trigger threshold, it means that the user's current geographic location abnormality and interest change frequency have reached a critical level that requires a system response. This means that the user's behavior pattern has changed significantly, and may have entered a new field of interest or shown new needs. In this case, it is necessary to immediately dynamically adjust the scope of the field of interest, recalculate the influence radius and center position to ensure that the field of interest can accurately cover the user's current activity area. At the same time, update the priority of the recommended content to ensure that the recommended content can better match the user's current interests and needs, thereby improving the accuracy of the recommendation and user experience.

[0112] After determining that the trigger value is greater than or equal to the trigger threshold, the interest field creation module and the position synchronization module are directly called to recalculate the influence radius and the center point and update the priority of the recommended content.

[0113] When updating the priority of recommended content, the weight of each recommended content is adjusted by comprehensively analyzing the latest distance between the user's current location and the center of the field of interest, behavioral feedback (such as click-through rate, dwell time), and the size of the trigger value. Specifically, the content with the highest matching degree with the user's current field of interest is selected, and its order in the recommendation list is adjusted according to the influence of the trigger value to ensure that the most relevant content is presented first. At the same time, the weight of content with which users show a high frequency of interaction is appropriately increased, and its display priority is increased to ensure that the recommended content is highly consistent with the actual needs of users.

[0114] The calculation formula for updating the priority of recommended content can be based on the comprehensive influence of the user's current location, behavioral feedback, and trigger value. For example, the following is a calculation formula with technical characteristics: ;

[0115] in:

[0116] : Time point Next content The updated weight (i.e. priority) of .

[0117] : Time point Next content The initial weight of .

[0118] : Position influence coefficient, which reflects the influence of the distance between the user's current position and the center of the field of interest on the priority.

[0119] : The minimum distance between the user's current position and the center of the field of interest, used to normalize the distance effect.

[0120] : Time point The user's current location and content The distance to the center of the relevant field of interest.

[0121] : Behavior feedback influence coefficient, reflecting the user's The impact of interactive behaviors (such as click-through rate and dwell time) on priority.

[0122] : Time point Next content Behavioral feedback score.

[0123] : Trigger value influence coefficient, reflecting the impact of the trigger value on the priority.

[0124] : Time point The trigger value calculated below.

[0125] This formula combines the distance between the user's current location and the field of interest, the content's behavioral feedback score, and the size of the trigger value. The closer the distance, the more frequent the user interaction, and the higher the trigger value, the higher the priority of the content.

[0126] Location affects Implementation, when the user is close to the center of the content-related field of interest, the distance factor increases and the priority is improved.

[0127] Behavioral feedback score Based on the user's interactive behavior, content with high feedback activity is prioritized.

[0128] Trigger value Used to capture global behavior changes and ensure that recommended content can be adjusted in real time when user interests or behaviors change significantly.

[0129] By using trigger values ​​constructed by location anomaly and interest switching frequency, the system can accurately identify significant changes in user behavior, thereby timely adjusting the scope of interest fields and the priority of recommended content. This ensures that the system can quickly provide the most relevant content when user needs change, improve the accuracy and personalization of recommendations, and ultimately improve user experience and satisfaction.

[0130] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0131] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0132] It should be noted that, in this article, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0133] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A multi-modal large-model active recommendation system based on LBS, characterized by: include: Interest field creation module, location synchronization module, interest fusion module and dynamic adjustment module; Interest field creation module: creates multiple interest fields, defines the center point, influence radius and weight of the interest field based on the user's long-term behavior data, covers the main activity areas around the city, and outputs the interest field information to the location synchronization module; Location synchronization module: synchronizes the user location and the state of the field of interest in real time, adjusts the influence range and priority of the field of interest to ensure consistency with the user's movement path, and outputs the synchronized state of the field of interest to the interest fusion module; Interest fusion module: integrates and analyzes short-term interests with current interest fields, dynamically adjusts interest field weights, balances the interactions of multiple interest fields, and outputs the fused interest field weights to the dynamic adjustment module; Dynamic adjustment module: It evaluates behavioral changes by combining the trigger value constructed by combining the abnormal degree of user's geographical location and the frequency of switching interest fields, dynamically adjusts the scope of interest field influence, recalculates the influence radius and center point, and updates the priority of recommended content; The operation process of the interest field creation module is as follows: S1.1, comprehensively collect users’ historical travel records; classify users’ preference types and summarize them into main areas of interest; S1.2, in the urban periphery, LBS technology is used to identify key geographical locations that users have visited many times in the past; S1.3, construct corresponding interest fields according to the user's preference type and the identified key geographical locations; each interest field is composed of a number of geographical location points, forming a basic structure of an interest field in geographical space; To ensure that the field of interest can accurately cover the user's activity area, the center point, influence radius and related weight parameters are defined for each field of interest during initialization; S1.4, in the process of constructing the field of interest, use multivariate regression analysis to calculate the comprehensive weight of user behavior data, and use weighted Voronoi diagram partitioning technology to optimize the space of the field of interest, and introduce environmental data into the graph structure model; according to the user's preference type and behavior pattern at each location, dynamically adjust the influence radius of each field of interest, and use the gravity model to calculate the interaction parameters between different fields of interest, so as to optimize the spatial layout of the field of interest; S1.5, after the creation and initialization of the field of interest, the path planning algorithm is used to simulate the path analysis of the user's recent behavior data; the center point and influence radius of the field of interest are corrected in combination with the geographical weighted regression model to ensure that the field of interest is highly consistent with the user's long-term behavior pattern; during the correction process, the user behavior change trend is predicted through time series analysis, and the Kalman filter algorithm is used to dynamically adjust the weight parameters; The operation process of the position synchronization module is as follows: S2.1, when the user enters the tourist area around the city, the user's current location is obtained in real time through LBS technology; the location data is expressed in the form of geographic coordinates; to ensure the accuracy of the data, the Kalman filter is used to filter the noise and smooth the trajectory of the location data to obtain the user's location; S2.2, after obtaining the user's current location, calculate the geographic distance between the user's location and the center point of each field of interest; use the Haversine formula in the earth coordinate system to calculate the distance to reflect the effect of the earth's curvature on the distance; S2.3, dynamically adjusting the influence radius of the field of interest according to the distance between the user's current position and the center point of the field of interest and the user's moving speed in the corresponding area; S2.4, based on the overlap between the user's position and the field of interest, the weight parameters of the field of interest are updated in real time, using the gravity model for calculation; S2.5, after completing the adjustment of the influence radius and weight parameters of the field of interest, apply the updated parameters to the state of the field of interest, and synchronize the coverage of the field of interest in real time; during the user's movement, continue to repeat the calculation steps of steps S2.1-S2.5 to ensure that the spatial layout and weight parameters of the field of interest are consistent with the actual position and behavior of the user; The operation process of the interest fusion module is as follows: S3.1, real-time acquisition of user behavior data, including recent social interactions and instant search behaviors; Recorded in timestamp format, social interaction data is , instant search data is ,in Indicates the time point of obtaining data; use time series analysis technology to smooth the behavior data, remove outliers, and obtain the trend function of the user's current short-term interest and ; S3.2 Converting users’ real-time behavior data into short-term interest influence coefficients , reflects the influence of short-term interests on the current interest field; the specific calculation formula is as follows: ;in, is the time decay coefficient, which is used to control the timeliness of short-term interests; Indicates a point in time in the past; and They are the trend functions of social interaction and instant search respectively; S3.3, based on the calculated short-term interest influence coefficient, perform a fusion analysis with the user's current interest field; let the weight of the user's current interest field be , then the weight of the fused field of interest is Dynamic adjustment is performed using the following formula: ;in, is the sensitivity parameter of the interest field, reflecting the responsiveness of the interest field to short-term interest; S3.4, while adjusting the weight of the current field of interest, balance the interactions between multiple fields of interest; use the interaction balance formula based on the gravity model: ;in, For interest field Field of interest The interaction forces between is the distance between the center points of the two fields of interest; and are the short-term interest influence coefficients of the current interest field and other interest fields respectively; S3.5, applying the interest field weights after fusion analysis to the recommendation algorithm, updating the recommended content in real time, and pushing the most relevant tourist destinations to the user’s current interest field; The operation process of the dynamic adjustment module is as follows: Position anomaly index generation logic: A1, collect the user's location data within a certain period of time and record it as a position vector sequence ,in and Represents the user at a point in time Longitude and latitude; standardize the location data and use normalization transformation to obtain the standardized position vector ; A2, based on the standardized position vector, construct the covariance matrix of the position data , which is used to describe the distribution characteristics of position data; the calculation formula of the covariance matrix is: ;in, is the total number of location data points; is the mean vector of the normalized position vector; represents the position vector after removing the mean; A3: Use Mahalanobis distance to measure the abnormality of the standardized position vector and calculate the abnormality score of each position point. , used to identify abnormal location behavior: ;in, is the mean vector of the normalized position vector; is the inverse matrix of the covariance matrix; Representing time point The position anomaly score of A4, based on the anomaly score at each time point, calculates the location anomaly index to comprehensively evaluate the degree of anomaly of user location data over a period of time; this is accomplished by accumulating anomaly scores and introducing a time decay factor , to reflect the timeliness of location anomalies: ;in, is the time decay factor, which controls the decay speed of the anomaly score; is the total time period for calculation; for a point in time in the past; Interest transition frequency generation logic: B1. First, collect the user's behavior data within a certain period of time. ,in Indicates a time point; the behavior data within a certain period of time is classified according to different areas of interest, and the behavior data in each area is represented as a vector ,in Indicates Areas of interest; B2, conduct time series analysis on the behavior data of each interest area to identify the changing trends of users' interests in different time periods; use discrete wavelet transform to decompose the time series data to obtain different frequency components ,in Indicates the level of the frequency component; ;in, is the wavelet basis function; for time point; is the total number of time points; B3, within each area of ​​interest, by calculating the interest transition rate at adjacent time points , identify the interest transition point, and the interest transition rate is: ;in, Indicate area of ​​interest At the point in time The transition rate of is the frequency component after wavelet decomposition; B4, for all interest areas, accumulate the frequency of interest transitions within a period of time to generate the interest transition frequency ; In the accumulation process, the time weighting factor is introduced , in order to control the timeliness of the transition; the calculation formula is: ;in, is the total number of areas of interest; is the observation period; for a point in time in the past; is the time weighting factor, which controls the timeliness of the transition.

2. The LBS-based multi-modal large model active recommendation system according to claim 1, characterized in that: The trigger value is obtained by comprehensively processing the location anomaly index and interest transition frequency using the weighted harmonic average method; if the trigger value is greater than or equal to the trigger threshold, it means that the user's current geographic location anomaly and interest change frequency have reached a critical level requiring a system response; this means that the user's behavior pattern has changed significantly. In this case, it is necessary to immediately and dynamically adjust the range of the field of interest, recalculate the influence radius and the center position to ensure that the field of interest can accurately cover the user's current activity area; at the same time, update the priority of the recommended content.

3. The LBS-based multi-modal large model active recommendation system according to claim 2, characterized in that: When updating the priority of recommended content, the weight of each recommended content is adjusted by comprehensively analyzing the latest distance between the user's current location and the center point of the field of interest, behavioral feedback, and the size of the trigger value.

Images