Intelligent service recommendation method based on multi-dimensional scene perception and dynamic portrait modeling

Abstract

The invention provides an intelligent service recommendation method based on multi-dimensional scene perception and dynamic portrait modeling. The intelligent service recommendation method comprises the steps that 1, distributed edge computing nodes are deployed, space-time tetrad data operated by a user are collected in real time, user behaviors and labels are stored, and a high-dimensional recommendation database is built; 2, constructing a dynamic portrait engine, and constructing a user long-term behavior pattern library; 3, deploying a real-time streaming and offline double-engine architecture, fusing real-time scene matching and offline portrait prediction, and performing personalized function and user deep demand analysis; step 4, for new users, synchronously calling geofences to obtain regional hot services, and realizing cold start optimization based on meta reinforcement learning in parallel; and 5, establishing an intelligent recommendation closed loop associated with weather characteristics. According to the method, multi-dimensional features such as space-time tetrad data, user tag information and dynamic interest weights are fully utilized, and the real-time performance, interpretability and generalization ability of a recommendation system can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and particularly relates to an intelligent service recommendation method based on multi-dimensional scenario perception and dynamic portrait modeling. Background Art

[0002] With the rapid development of the mobile Internet and the Internet of Things, recommendation systems play an important role in fields such as personalized services and meteorological services, providing accurate recommendations by analyzing user behavior and context data. However, traditional recommendation systems have significant limitations in dealing with real-time multi-dimensional data, dynamic user portraits, cold start problems, and meteorological feature correlation recommendations.

[0003] First of all, existing recommendation systems usually rely on centralized data processing, making it difficult to collect and process spatio-temporal data of user operations in real time, resulting in the inability to build a high-dimensional recommendation database to support accurate recommendations in dynamic scenarios. In terms of user portrait modeling, traditional methods are mostly based on static data, lacking a dynamic update mechanism for user interests, and it is difficult to capture long-term associations in behavior sequences, limiting the accuracy of personalized recommendations. Secondly, the architecture of recommendation systems often relies on a single processing engine, failing to effectively integrate real-time scenario matching and offline prediction, and it is difficult to balance the analysis of immediate needs and deep needs. In addition, for new users, the cold start problem is prominent. Existing systems usually rely on general popular recommendations and lack the ability to quickly adapt to user behavior, resulting in poor recommendation effects. Finally, in the field of meteorological service recommendations, existing systems rarely integrate weather feature data with user needs, and it is difficult to achieve automated and accurate meteorological service recommendations through semantic parsing and threshold alarms. Summary of the Invention

[0004] Object of the Invention: The technical problem to be solved by the present invention is to provide an intelligent service recommendation method based on multi-dimensional scenario perception and dynamic portrait modeling in view of the deficiencies of the prior art, including the following steps: Step 1, deploy distributed edge computing nodes, collect spatio-temporal quadruple data of user operations in real time, store user behavior and tags, and build a high-dimensional recommendation database; Step 2, build a dynamic portrait engine, calculate interest weights using the attention mechanism, and mine long-term associations in behavior sequences through Transformer (a neural network architecture based on the attention mechanism, which generates long-term behavior patterns for personalized recommendations by mining long-term associations in user behavior sequences), and build a user long-term behavior pattern library; Step 3, deploy a real-time stream plus offline dual-engine architecture, integrate real-time scenario matching and offline portrait prediction, and perform personalized function and user deep need analysis; Step 4, for new users, synchronously call geofencing to obtain regional popular services, and parallelly implement cold start optimization based on meta-reinforcement learning; Step 5, establish an intelligent recommendation closed-loop associated with weather features: By semantically parsing the keywords and threshold alarm information in meteorological service materials, automatically match and trigger relevant meteorological service recommendations.

[0005] Step 1 includes: Deploy edge computing nodes, and through GPS and WiFi fingerprint dual-mode positioning and meteorological API (Application Programming Interface, abbreviated as API), generate spatio-temporal quadruples in real time , where is the absolute time when the i-th user operation occurs, is the longitude value of the i-th user in the WGS-84 coordinate system at [[ID=ll]] time, [[ID=1:3]] is the latitude value of the i-th user in the WGS-84 coordinate system at time, refers to the normalized meteorological state of the i-th user at time; N represents the total number of data points in the set.

[0006] Step 1 also includes: Fusing the GPS and WiFi fingerprint positioning results through Kalman filtering: , , where is the finally adopted user location, and are the original longitude and latitude coordinates provided by the GPS module and the original longitude and latitude coordinates provided by the WiFi module respectively, is the positioning source variance of the WiFi module, is the positioning source variance of the GPS module, is the dynamically calculated weight coefficient; Map the weather state to a 6-dimensional vector: temperature, humidity, wind speed, precipitation, PM2.5, ultraviolet rays, and normalize it using a piecewise linear function: , where ; refers to the normalized value of the k-th meteorological parameter, refers to the original observed value of the k-th meteorological parameter, and are the historical mean and standard deviation respectively; Adopt an improved spatial clustering of applications with noise based on spatio-temporal density ST-DBSCAN algorithm, and define a spatio-temporal composite distance: , wherein, denotes the i-th point and the j-th point the spatio-temporal composite distance therebetween, denotes the geographical distance, is the time normalization coefficient, is the spatial weight factor, , are the i-th time scalar and the j-th time scalar respectively; If , is the clustering parameter, then it is determined that the i-th point and the j-th point belong to the same cluster, completing the spatio-temporal clustering division to obtain the spatio-temporal clustering result; Embed the user label into a tuple and expand it into a five-tuple form: , wherein is the user label of the i-th user at time , and the following formula is used for label weight calculation: , wherein, is the dynamic weight of the user for label m at time , is the time interval since the last update, represents the user behavior o, is the weight score of behavior o, () is the indicator function, is the set of associated behaviors, and e is the natural constant; Finally, integrating the spatio-temporal quadruple and user label information, the construction of a high-dimensional recommendation database is realized.

[0007] Step 2 includes: calculating the dynamic interest intensity of the user for different service categories by using spatio-temporal encoding and attention mechanism, introducing a time decay factor, and the updated interest weight calculation formula is: , wherein, represents the normalized interest weight of the user for label m at time step t, is the query vector, is the projection matrix of label m, is the scaling factor, is the key matrix, that is, the semantic information encoding historical behaviors; Perform time decay correction calculation: , wherein, Denote the final interest weight of user for tag m at time t. is the decay rate; Complete the aggregation of user behavior habits and tags according to the final interest weight. Input the five-tuple in step 1 into the Transformer-XL (Transformer with eXtra Long context) with an ultra-long context. The Transformer-XL includes a recurrent memory module and relative position encoding to identify the long-term behavior patterns of the user. The recurrent memory module is responsible for caching the hidden states of the previous segment (the "previous segment" refers to when processing long sequence data, the input sequence is divided into multiple fixed-length segments, and the "previous segment" is the segment before the currently processed sequence segment, which is a standard concept in the Transformer-XL architecture) and completing the splicing of historical memories during the calculation of the current segment. The hidden state of the previous segment is expressed as: , Splice the historical memory during the calculation of the current segment: , where is the memory cache of the previous segment, which consists of the hidden states of each layer of the previous segment ; is the sequence length, represents the L-th hidden state in the previous segment; is the attention output of the current segment, , , are the query, key, and value matrices of the current segment respectively; represents the attention calculation; is the key matrix of the previous segment; is the value matrix of the previous segment.

[0008] Step 3 includes: Step 3-1, scene matching of the real-time stream engine: Query the spatio-temporal clustering results generated by the improved ST-DBSCAN algorithm in step 1 according to the finally adopted user location to find the clustering cluster to which the user's current scene belongs ; Calculate the similarity between the user's current scene and historical scenes , and the formula is: , where is the geographical distance, is the time interval, is the clustering cluster of the historical scene; Step 3-2, Dynamic Interest Weight Adjustment of the Real-Time Streaming Engine: Adjust the interest weights of users for different service categories according to the real-time scenario matching results , the formula is: , Generate a real-time recommendation service list according to the adjusted interest weights ; Step 3-3, Deep Demand Prediction of the Offline Engine: Use the user's long-term behavior pattern library to predict the user's potential needs in the current scenario, and calculate the weights of the user's deep needs through the attention mechanism : , where is the similarity score of the user for demand γ, is the similarity score of the user for the historical scenario δ, and H is the total number of historical scenarios; Step 3-4, Demand Aggregation of the Offline Engine: Aggregate the user's historical behavior and label information according to the deep demand weights to generate a deep demand prediction vector : , where represents the feature representation vector of the user's historical behavior and label information in the historical scenario; Step 3-5, Fusion of the Real-Time Streaming and Offline Engines: Dynamically adjust the weights of the real-time streaming and offline engines according to the activity level and time interval of the user's behavior , and then generate a recommendation service list according to the final recommendation score : , , where is the adjustment parameter, is the real-time recommendation score, is the offline recommendation score.

[0009] Step 4 includes: Query all services within the geofence according to the user's current location , obtain the access frequency of the services, and perform popularity ranking according to the weight of each service , and generate a list of popular services: , where , are weight parameters, , are the access frequency and rating of service s respectively; Use the pre-trained meta-policy network to adapt to the behavior patterns of new users. Define the user state by the user's current geographical location, time, and the list of services the user has interacted with, and define the reward function r: , where indicates whether the user clicks on the recommended service , is the residence time of the user on service , , are weight parameters; Finally, use meta-reinforcement learning to update the policy network parameters and optimize the recommendation policy.

[0010] In step 4, the update formula for the policy network parameters is: , where are the pre-trained meta-policy network parameters, is the learning rate, is the gradient of the reward, represents the value of the reward function at time step t.

[0011] Step 5 includes: performing text analysis on meteorological service materials, extracting keywords from them, and building a meteorological database; at the same time, defining threshold rules, which are expressed as a set , where is the x-th meteorological element, is the corresponding threshold; When keywords or meteorological elements are matched and exceed the threshold, the recommendation mechanism is automatically triggered to generate a list of meteorological services closely related to the user's current environment : , where is the keyword matching set, is the threshold alarm information set; Construct a high-frequency word library in the meteorological database by performing text statistical analysis on meteorological service materials, and match the current environmental characteristics with keywords in real time to generate a keyword matching set containing relevant services; Perform text analysis on meteorological service materials through natural language processing technology, including text cleaning, word segmentation, and part-of-speech tagging. Combine TF-IDF and TextRank algorithms to extract keywords, supplemented by BERT word embedding for semantic expansion, and construct a structured meteorological database keyword library containing keywords, categories, weights, semantic vectors, and associated services. During real-time matching, use the spatio-temporal quadruple in step 1 to extract the current environmental features, compare the environmental feature tags with the keyword library through exact matching and semantic similarity calculation, and combine the threshold rule to preferentially trigger high-weight keywords to generate a keyword matching set. , and map it to relevant meteorological services to form a real-time, accurate, and diverse recommended service list. Based on a predefined set of threshold rules , monitor real-time meteorological data. When meteorological elements exceed the threshold , trigger the corresponding service to generate a threshold alarm information set.

[0012] The present invention also provides an electronic device, including a processor and a memory. The memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0013] The present invention also provides a storage medium storing a computer program or instruction, and when the computer program or instruction runs on a computer, it executes the steps of the method.

[0014] Beneficial effects: The method of the present invention performs fusion through a dynamic weight allocation mechanism, dynamically adjusts the weights of the real-time stream and the offline engine according to the activity level and time interval of user behavior, and thus generates the final recommendation service. This method makes full use of multi-dimensional features such as spatio-temporal quadruple data, user tag information, and dynamic interest weights, and can effectively improve the real-time performance, interpretability, and generalization ability of the recommendation system. In addition, by introducing multi-modal data fusion and diversity constraints, the diversity of the recommendation results and the user experience are further enhanced. The advantage of this technology is that it can simultaneously meet the immediate and deep needs of users, provide accurate and diverse recommendation services, and at the same time has strong adaptability and scalability, and is applicable to complex dynamic scenarios. The advantage of the solution of the present invention is that it can provide a fast and accurate initial recommendation experience for new users. The geofencing service ensures the geographical relevance of the recommendation, while the meta-reinforcement learning improves the personalization degree of the recommendation by quickly learning user preferences. This method combining geofencing and meta-reinforcement learning not only solves the cold start problem, but also enhances the adaptability and scalability of the recommendation system, and is applicable to a dynamically changing user environment. The present invention can respond to weather changes in real time, provide accurate and timely meteorological service recommendations, and ensure that users obtain reliable information support under dynamic weather conditions. Keyword matching ensures the accuracy of the recommended content, and the threshold information prompt provides timely warnings, significantly improving the user experience and the practicality of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The following further specifically describes the present invention in conjunction with the drawings and specific embodiments, and the above and / or other advantages of the present invention will become clearer.

[0016] Figure 1 is the flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] As Figure 1 shown, an embodiment of the present invention provides an intelligent service recommendation method based on multi-dimensional scenario perception and dynamic portrait modeling, including: Step 1, deploy distributed edge computing nodes, and generate spatio-temporal quadruples in real time through GPS / WiFi fingerprint dual-mode positioning and meteorological API , where records the absolute time when the user operation occurs, is the longitude value in the WGS-84 coordinate system, is the latitude value in the WGS-84 coordinate system, refers to the normalized meteorological state.

[0018] Fuse the GPS and WiFi fingerprint positioning results through Kalman filtering: , , where is the finally adopted user location, and are the original latitude and longitude coordinates provided by the GPS module and the original latitude and longitude coordinates provided by the WiFi module respectively, is the positioning source variance of the WiFi module, is the positioning source variance of the GPS module, is the dynamically calculated weight coefficient; Map the weather status to a 6-dimensional vector (temperature / humidity / wind speed / precipitation / PM2.5 / ultraviolet), and normalize it using a piecewise linear function: , where ; refers to the normalized value of the kth meteorological parameter (the 1st meteorological parameter is temperature, the 2nd meteorological parameter is humidity, the 3rd meteorological parameter is wind speed, the 4th meteorological parameter is precipitation, the 5th meteorological parameter is PM2.5, and the 6th meteorological parameter is ultraviolet), refers to the original observed value of the kth meteorological parameter, and are the historical mean and standard deviation respectively.

[0019] Adopt the improved spatio-temporal density-based spatial clustering of applications with noise (ST-DBSCAN) algorithm and define the spatio-temporal composite distance: , where refers to the spatio-temporal composite distance between the ith point and the jth point , refers to the geographical distance, is the time normalization coefficient, is the spatial weight factor, and are the ith time scalar and the jth time scalar respectively; If , is the clustering parameter, then it is judged that the ith point and the jth point belong to the same cluster, complete the spatio-temporal clustering division, and obtain the spatio-temporal clustering result; Embed the user label into a tuple and expand it into a five-tuple form: , where is the user tag of the i-th user at time, and the tag weight is calculated using the following formula: , where is the dynamic weight of the user for category m at time , is the time interval since the last update, is the weight score of behavior o, (·) is an indicator function used to determine whether a certain action belongs to the set . If it belongs, the value is 1, otherwise the value is 0; is the set of associated behaviors.

[0020] Based on the above spatio-temporal data and user tag information, the construction of a high-dimensional recommendation database is realized.

[0021] Step 2 proposes a method for constructing a user portrait based on dynamic weight calculation and long-term dependence modeling. The spatio-temporal aware multi-head attention mechanism is used to quantify the user interest weight, and the improved Transformer-XL architecture is used to capture the cross-period behavior pattern.

[0022] The system uses spatio-temporal encoding and attention mechanism to calculate the dynamic interest intensity of the user for different service categories, and introduces a time decay factor to achieve accurate modeling of interest evolution. The updated interest weight calculation: , where represents the normalized interest weight of the user for tag m at time step t, is the query vector, is the projection matrix of category m, is the scaling factor, is the key matrix, that is, the semantic information encoding historical behaviors; Perform time decay correction calculation: , In the formula, represents the final interest weight of the user for tag m at time t, is the decay rate, Δt is the time interval, and the longer the time interval, the more significant the weight decay of historical interests. According to the final interest weight, the aggregation of user behavior habits and tags is completed; The input sequence for long-term dependency modeling comes directly from the multi-tuple in step 1. Through the recurrent memory module and relative position encoding of Transformer-XL, it effectively identifies the long-term behavior patterns of users and solves the bottleneck of traditional models in modeling long sequences.

[0023] The hidden state of the previous segment is expressed as: , When calculating the current segment, join the historical memory: , in, It is the memory cache of the previous section, which is composed of the hidden states of each layer in the previous section. composition, is the sequence length, Indicates the previous period (time step ) in the Lth hidden state; is the attention output of the current segment, 、 、 are the query, key, and value matrices of the current segment respectively; represents attention calculation; is the key matrix of the previous segment; is the value matrix of the previous section. The modeling results can obtain the user's long-term behavior pattern library and behavior prediction vector, which serve as the input data of the recommendation model.

[0024] Step 3: Deploy a real-time streaming and offline dual-engine architecture, integrating real-time scene matching with offline profile prediction to perform personalized functions and in-depth user needs analysis. This includes the following steps: Step 3-1, scene matching of real-time streaming engine: based on the final user location Query the spatiotemporal clustering results generated in step 1 to find the cluster to which the user's current scene belongs . Calculate the similarity between the user's current scene and historical scenes , the formula is: , in, is the geographical distance, is the time interval, is the cluster of historical scenes.

[0025] Step 3-2, dynamic interest weight adjustment of real-time streaming engine: adjust the user's interest weight for different service categories based on the real-time scene matching results , the formula is: , Based on the adjusted interest weight , generate a list of real-time recommendation services; Step 3-3, deep demand prediction of the offline engine: Utilize the user's long-term behavior pattern library to predict the user's potential demands in the current scenario, and calculate the weights of the user's deep demands through the attention mechanism : , where is the similarity score of the user for demand γ, is the similarity score of the user for historical scenario δ, and H is the total number of historical scenarios; Step 3-4, demand aggregation of the offline engine: Aggregate the user's historical behavior and label information according to the deep demand weights to generate a deep demand prediction vector : , Step 3-5, fusion of the real-time stream and the offline engine: Dynamically adjust the weights of the real-time stream and the offline engine according to the activity level and time interval of the user's behavior , and then according to the final recommendation score , generate a list of recommended services: , , where, is the adjustment parameter, Δt is the time interval of the user's last interaction, is the real-time recommendation score, is the offline recommendation score.

[0026] Step 4, for new users, synchronously call the geofence to obtain the popular services in the area, and parallelly implement cold start optimization based on meta-reinforcement learning; including: According to the user's current location , query all services within the geofence to obtain information such as the access frequency of the services. According to the weight of each service , perform popularity ranking and generate a list of popular services: , where, , are weight parameters used to balance the influence of access frequency and rating, , are respectively the access frequency and rating of service s.

[0027] Based on meta-reinforcement learning, the system can quickly adapt to the behavior patterns of new users and optimize the recommendation strategy. The pre-trained meta-policy network is used to quickly adapt to the behavior patterns of new users, and the user state is defined through the user's current geographical location, time, and the list of services the user has interacted with. Define the reward function , which is calculated based on the user's interaction behavior: , where represents whether the user clicks on the recommended service ( represents the click situation of the user using the service, the value of which will be stored in the database), is the residence time of the user on service , , are weight parameters; Finally, use meta-reinforcement learning to update the policy network parameters , and optimize the recommendation strategy. The formula is: , where are the parameters of the pre-trained meta-policy network, is the learning rate, is the gradient of the reward, represents the value of the reward function at time step t, which is used to quantify the performance of the recommendation strategy in the current state.

[0028] Step 5, establish an intelligent recommendation closed-loop associated with weather features: By semantically parsing the keywords and threshold alarm information in the meteorological service materials, automatically match and trigger relevant meteorological service recommendations, including: performing text analysis on the meteorological service materials, extracting the keywords therein, and building a meteorological database; at the same time, defining threshold rules, which are represented as a set , where is the x-th meteorological element, is the corresponding threshold, and the value of the threshold is set based on the statistical analysis of historical meteorological data and expert experience. The specific method is: First, collect historical observation data related to the meteorological element (including temperature, humidity, wind speed, precipitation, PM2.5, ultraviolet rays, etc.), and calculate its mean and standard deviation ; then, according to the business requirements of the meteorological service and the safety warning standard, set the threshold , where k is the adjustment coefficient, usually ranging from 1.5 to 3, which is determined by meteorological experts according to the fluctuation characteristics of specific meteorological elements and application scenarios; for extreme weather scenarios, calibration is carried out in combination with the standard thresholds issued by the National Meteorological Administration to ensure that can effectively trigger relevant service recommendations; When keywords or meteorological elements are matched and exceed the threshold, the recommendation mechanism is automatically triggered to generate a list of meteorological services closely related to the user's current environment : , Among them, is the keyword matching set, is the threshold alarm information set; By performing text statistical analysis on meteorological service materials in the past 10 years, a high-frequency word library in the meteorological database is constructed to match the current environmental characteristics with keywords in real time and generate a keyword matching set containing relevant services; Through natural language processing technology, text analysis of meteorological service materials is carried out, including text cleaning, word segmentation, part-of-speech tagging, and keywords are extracted by combining TF-IDF and TextRank algorithms, supplemented by BERT word embedding for semantic expansion, to construct a structured meteorological database keyword library containing keywords, categories, weights, semantic vectors and associated services, supporting regular updates and dynamic optimization; during real-time matching, the current environmental characteristics are extracted using the spatio-temporal quadruple in step 1, and the environmental feature tags are compared with the keyword library through exact matching and semantic similarity calculation, and high-weight keywords are preferentially triggered in combination with the threshold rule to generate a keyword matching set , and mapped to relevant meteorological services to form a real-time, accurate and diverse recommendation service list;

[0029] Based on a predefined set of threshold rules , monitor real-time meteorological data. When the meteorological element exceeds the threshold , trigger the corresponding service and generate a threshold alarm information set.

[0030] To improve the pertinence of meteorological services, this embodiment has developed a personalized recommendation technology route based on user behavior and weather characteristics. A dynamic recommendation technology driven by user behavior. Based on user interaction logs, key information such as access time, geographical location, product preferences, and weather query records are extracted to construct a three-dimensional feature database of time-space-behavior. On this basis, by using high-frequency behavior pattern mining and interest weight calculation algorithms, the user tags are dynamically updated every hour, thus constructing a real-time updated user portrait. The system adopts a dual-engine architecture of "real-time stream computing + offline modeling". The real-time module makes product recommendations immediately according to the user's current situation; the offline module optimizes the user's long-term preference model through in-depth analysis of the user's historical behavior. For new users, the system provides default recommendations based on role presets, and quickly constructs a short-term portrait using the initial interaction data in the background, effectively alleviating the cold start problem.

[0031] The present invention provides an intelligent service recommendation method based on multi-dimensional scenario perception and dynamic portrait modeling. There are many methods and ways to specifically implement this technical solution. The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. Each component not clearly defined in this embodiment can be implemented using existing technologies.

Claims

1. An intelligent service recommendation method based on multi-dimensional scenario perception and dynamic portrait modeling, characterized in that, It includes the following steps: Step 1: Deploy distributed edge computing nodes, collect spatio-temporal quadruple data of user operations in real time, store user behaviors and labels, and build a high-dimensional recommendation database; Step 2: Build a dynamic portrait engine, calculate interest weights using the attention mechanism, and mine long-term associations of behavior sequences through Transformer to build a user long-term behavior pattern library; Step 3: Deploy a real-time stream plus offline dual-engine architecture, fuse real-time scene matching and offline portrait prediction, and conduct personalized function and user deep demand analysis; Step 4: For new users, synchronously call the geofence to obtain popular services in the area, and parallelly implement cold start optimization based on meta-reinforcement learning; Step 5: Establish an intelligent recommendation closed-loop associated with weather features: Automatically match and trigger relevant weather service recommendations by semantic parsing of keywords and threshold alarm information in meteorological service materials.

2. The method according to claim 1, wherein Step 1 includes: Deploying edge computing nodes to generate spatio-temporal quadruples in real time through GPS, WiFi fingerprint dual-mode positioning, and meteorological APIs. , where is the absolute time when the i-th user operation occurs, is the longitude value of the i-th user in the WGS-84 coordinate system at time, is the latitude value of the i-th user in the WGS-84 coordinate system at time, refers to the normalized meteorological state of the i-th user at time; N represents the total number of data points in the set.

3. The method according to claim 2, characterized in that, Step 1 further includes: Fusing the GPS and WiFi fingerprint positioning results through Kalman filtering: ， ， Among them is the finally adopted user location, , are the original longitude and latitude coordinates provided by the GPS module and the original longitude and latitude coordinates provided by the WiFi module respectively, is the positioning source variance of the WiFi module, is the positioning source variance of the GPS module, is the dynamically calculated weight coefficient; Map the weather conditions to a 6-dimensional vector: temperature, humidity, wind speed, precipitation, PM2.5, and ultraviolet radiation, and normalize using a piecewise linear function: ， wherein ; refers to the normalized value of the k-th meteorological parameter, refers to the original observed value of the k-th meteorological parameter, 、 are the historical mean and standard deviation respectively; Adopt an improved density-based spatial clustering of applications with noise in space-time ST-DBSCAN algorithm to define the spatio-temporal composite distance: ， Among them, refers to the i-th point and the j-th point The spatio-temporal composite distance between them refers to the geographical distance is the time normalization coefficient is the spatial weight factor 、 are the i-th time scalar and the j-th time scalar respectively; If , is the clustering parameter, then it is determined that the \(i\)-th point and the \(j\)-th point belong to the same cluster, completing the spatio-temporal clustering division and obtaining the spatio-temporal clustering result; Embed the user tag into a tuple and expand it into a five-tuple form: , where is the user tag of the i-th user at time . The tag weight is calculated using the following formula: ， wherein, is the dynamic weight of label m for the user at time , is the time interval since the last update, represents user behavior o, is the weight score of behavior o, I(·) is the indicator function, is the set of associated behaviors, and e is the natural constant; Finally, comprehensively considering the spatio-temporal quadruple and user label information, realize the construction of the high-dimensional recommendation database.

4. The method according to claim 3, wherein Step 2 includes: Using spatio-temporal encoding and the attention mechanism to calculate the dynamic interest intensity of users for different service categories, introducing a time decay factor, and the updated interest weight calculation formula is: ， Among them, represents the normalized interest weight of user for label m at time step t, is the query vector, is the projection matrix of label m, is the scaling factor, is the key matrix; Conduct time decay correction calculation: ， Among them, represents the user's final interest weight for label m at time , and is the decay rate; According to the final interest weights, complete the aggregation of user behavior habits and labels; Input the quintuple of Step 1 into the Transformer-XL with ultra-long context, where the Transformer-XL with ultra-long context includes a recurrent memory module and relative position encoding, identify the long-term behavior patterns of users, and the recurrent memory module is responsible for caching the hidden state of the previous segment and completing the splicing of historical memories during the calculation of the current segment; The hidden state of the previous segment is expressed as: ， Splice historical memories during the calculation of the current segment: ， Among them, is the memory cache of the previous segment, which consists of the hidden states of each layer of the previous segment ; is the sequence length, indicating the L-th hidden state in the previous segment; is the attention output of the current segment, , , are the query, key, and value matrices of the current segment respectively; represents the attention calculation; is the key matrix of the previous segment; is the value matrix of the previous segment.

5. The method according to claim 4, wherein Step 3 includes: Step 3-1, scene matching of the real-time stream engine: According to the finally adopted user location Query the spatio-temporal clustering results generated by improving the spatio-temporal density-based noise application spatial clustering ST-DBSCAN algorithm in Step 1 to find the clustering cluster to which the user's current scene belongs ; Calculate the similarity between the user's current scene and historical scenes , and the formula is: ， wherein is the geographical distance, is the time interval, is the clustering cluster of historical scenarios; Step 3-2, dynamic interest weight adjustment of the real-time streaming engine: Adjust the user's interest weights for different service categories according to the real-time scenario matching results , and the formula is: ， Based on the adjusted interest weights , generate a real-time recommendation service list; Step 3-3, Deep Demand Prediction of the Offline Engine: Utilize the user's long-term behavior pattern library to predict the user's potential demands in the current scenario, and calculate the weights of the user's deep demands through the attention mechanism : ， where is the similarity score of the user to requirement γ, is the similarity score of the user to historical scenario δ, and H is the total number of historical scenarios; Step 3-4, Requirement Aggregation of the Offline Engine: Aggregate the historical behavior and tag information of the user according to the deep requirement weights to generate a deep requirement prediction vector : ， Among them represents the feature representation vector of the user's historical behavior and label information in the historical scenario; Step 3-5, Integration of Real-time Stream and Offline Engine: Dynamically adjust the weights of the real-time stream and the offline engine according to the activity level and time interval of user behavior , and then based on the final recommendation score , generate a list of recommendation services: ， ， Among them, is an adjustment parameter, is the real-time recommendation score, is the offline recommendation score.

6. The method according to claim 5, wherein Step 4 includes: Based on the user's current location , query all services within the geofence, obtain the access frequency of the services, and according to the weight of each service , perform a popularity ranking and generate a list of popular services: ， Among them, , are weight parameters, , are the access frequency and score of service s, respectively; Use a pre-trained meta-policy network to adapt to the behavior patterns of new users, complete the definition of user states through the user's current geographical location, time, and the list of services the user has interacted with, and define the reward function r: ， Among them, indicates whether the user clicks on the recommended service , is the residence time of the user on the service , , are weight parameters; Finally, use meta-reinforcement learning to update the policy network parameters , and optimize the recommendation policy.

7. The method according to claim 6, characterized in that, In step 4, the update formula for the policy network parameters is as follows: ， Among them, are the parameters of the pre-trained meta-policy network, is the learning rate, is the gradient of the reward, represents the value of the reward function at time step t.

8. The method according to claim 7, characterized in that Step 5 includes: performing text analysis on meteorological service materials, extracting keywords therein, and building a meteorological database; meanwhile, defining threshold rules, and the threshold rules are expressed as a set , where is the x-th meteorological element, is the corresponding threshold; When a keyword or meteorological element is matched and exceeds the threshold, the recommendation mechanism is automatically triggered to generate a list of meteorological services closely related to the user's current environment : ， Among them, is a keyword matching set, is a threshold alarm information set; Through text statistical analysis of meteorological service materials, build a high-frequency word library in the meteorological database, match the current environmental features with keywords in real time, and generate a keyword matching set containing relevant services; Perform text analysis on meteorological service materials through natural language processing technology, including text cleaning, word segmentation, and part-of-speech tagging. Combine TF-IDF and TextRank algorithms to extract keywords, supplemented by BERT word embeddings for semantic expansion, and construct a structured meteorological database keyword library containing keywords, categories, weights, semantic vectors, and associated services. During real-time matching, use the spatio-temporal quadruple in Step 1 to extract the current environmental features, compare the environmental feature tags with the keyword library through exact matching and semantic similarity calculation, and combine threshold rules to preferentially trigger high-weight keywords to generate a keyword matching set and map it to relevant meteorological services to form a real-time, accurate, and diverse recommended service list; Based on a predefined set of threshold rules , monitor real-time meteorological data, and when meteorological elements exceed the threshold , trigger the corresponding service to generate a set of threshold alarm messages.

9. An electronic device, characterized in that, It includes a processor and a memory, and the memory stores program code. When the program code is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 8.

10. A storage medium, characterized in that, Store a computer program or instruction. When the computer program or instruction runs on a computer, it executes the steps of the method according to any one of claims 1 to 8.

Images