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

Abstract

This invention provides an intelligent service recommendation method based on multi-dimensional scene perception and dynamic profile modeling, including: Step 1: Deploy distributed edge computing nodes to collect spatiotemporal quadruple data of user operations in real time, store user behaviors and tags, and build a high-dimensional recommendation database; Step 2: Build a dynamic profile engine and construct a library of long-term user behavior patterns; Step 3: Deploy a real-time streaming and offline dual-engine architecture that integrates real-time scene matching with offline profile prediction to perform personalized functions and analyze in-depth user needs; Step 4: For new users, synchronously invoke geofencing to obtain regional popular services and implement cold start optimization based on meta-reinforcement learning; Step 5: Establish an intelligent recommendation closed loop associated with weather features. This method fully utilizes multi-dimensional features such as spatiotemporal quadruple data, user tag information, and dynamic interest weights, effectively improving the real-time performance, interpretability, and generalization capabilities of the recommendation system.

Description

Technical Field

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

[0002] With the rapid development of mobile internet and the Internet of Things, recommendation systems are playing a vital role in areas such as personalized services and weather services. They provide precise recommendations by analyzing user behavior and contextual data. However, traditional recommendation systems have significant limitations in processing real-time multidimensional data, dynamic user profiles, cold start issues, and weather-related recommendations.

[0003] First, existing recommendation systems typically rely on centralized data processing, making it difficult to collect and process spatiotemporal data on user actions in real time. This makes it impossible to build high-dimensional recommendation databases to support accurate recommendations in dynamic scenarios. Regarding user profile modeling, traditional methods are mostly based on static data, lacking a dynamic update mechanism for user interests. This makes it difficult to capture long-term correlations in behavioral sequences, limiting the accuracy of personalized recommendations. Second, the architecture of recommendation systems often relies on a single processing engine, failing to effectively integrate real-time scene matching with offline predictions, making it difficult to balance the analysis of immediate needs with underlying needs. Furthermore, for new users, the cold start problem is prominent. Existing systems typically rely on general popular recommendations, lacking the ability to quickly adapt to user behavior, resulting in poor recommendation results. Finally, in the field of meteorological service recommendations, existing systems rarely integrate weather characteristic data with user needs, making it difficult to achieve automated and accurate meteorological service recommendations through semantic parsing and threshold alarms. Summary of the Invention

[0004] Purpose of the invention: The technical problem to be solved by the present invention is to address the deficiencies of the existing technology and provide an intelligent service recommendation method based on multi-dimensional scene perception and dynamic portrait modeling, comprising the following steps:

[0005] Step 1: Deploy distributed edge computing nodes to collect spatiotemporal quadruple data of user operations in real time, store user behaviors and tags, and build a high-dimensional recommendation database.

[0006] Step 2: Build a dynamic profiling engine, use the attention mechanism to calculate interest weights, and use the Transformer (a neural network architecture based on the attention mechanism that mines long-term correlations in user behavior sequences to generate long-term behavior patterns for personalized recommendations) to mine long-term correlations in behavior sequences and build a library of long-term user behavior patterns.

[0007] 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.

[0008] Step 4: For new users, geo-fencing is synchronously called to obtain popular services in the area, and cold start optimization based on meta-reinforcement learning is implemented in parallel;

[0009] Step 5: Establish an intelligent recommendation closed loop associated with weather characteristics: automatically match and trigger relevant weather service recommendations by semantically analyzing keywords and threshold alarm information in meteorological service materials.

[0010] Step 1 includes: deploying edge computing nodes, generating spatiotemporal quadruple in real time through GPS, WiFi fingerprint dual-mode positioning and meteorological API (Application Programming Interface, API for short) ,in It is the absolute time when the i-th user operation occurs. is the i-th user in the WGS-84 coordinate system The longitude value of the time, is the i-th user in the WGS-84 coordinate system The latitude value of the time, Refers to the i-th user in The normalized meteorological state at the time; N represents the total number of data points in the set.

[0011] Step 1 also includes: fusing the GPS and WiFi fingerprint positioning results through Kalman filtering:

[0012] ,

[0013] ,

[0014] in is the final user location, 、 They are the original longitude and latitude coordinates provided by the GPS module and the original longitude and latitude coordinates provided by the WiFi module. is the positioning source variance of the WiFi module, is the positioning source variance of the GPS module, To dynamically calculate the weight coefficient;

[0015] The weather status Mapped into a 6-dimensional vector: temperature, humidity, wind speed, precipitation, PM2.5, and ultraviolet light, and normalized using a piecewise linear function:

[0016] ,

[0017] in ; refers to the normalized value of the kth meteorological parameter, refers to the original observation value of the kth meteorological parameter, 、 are the historical mean and standard deviation respectively;

[0018] The improved spatial clustering ST-DBSCAN algorithm based on noise application of spatiotemporal density is used to define the spatiotemporal composite distance:

[0019] ,

[0020] in, Refers to the i-th point and the jth point The space-time composite distance between Refers to 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;

[0021] if , Is the clustering parameter, then determine the i-th point and the jth point Belong to the same cluster, complete the division of spatiotemporal clustering, and obtain the spatiotemporal clustering results;

[0022] Embed the user tag into the tuple and expand it into a five-tuple form: ,in For the i-th user at time The user tag is calculated using the following formula:

[0023] ,

[0024] in, For users at all times Dynamic weight for label m, is the time interval since the last update, Represents user behavior o, is the weight score of behavior o, (·) is the indicator function, is a set of associated behaviors, e is a natural constant;

[0025] Finally, the spatiotemporal quadruple and user tag information are integrated to build a high-dimensional recommendation database.

[0026] Step 2 includes: using spatiotemporal coding and attention mechanism to calculate the dynamic interest intensity of users in different service categories, introducing a time decay factor, and the updated interest weight calculation formula is:

[0027] ,

[0028] in, represents the normalized interest weight of the user on 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, which encodes the semantic information of historical behavior;

[0029] Perform time decay correction calculations:

[0030] ,

[0031] in, represents the user’s final interest weight for tag m at time t, is the decay rate;

[0032] Based on the final interest weight, complete the aggregation of user behavior habits and tags;

[0033] The quintuple from step 1 is fed into the Transformer-XL (Transformer with eXtra Long Context). The Transformer-XL includes a recurrent memory module and relative position encoding to identify long-term user behavior patterns. The recurrent memory module caches the hidden state of the previous segment ("previous segment" refers to the segment before the current segment when processing long sequence data, where the "previous segment" is the segment before the currently processed segment, a standard concept in the Transformer-XL architecture) and completes historical memory splicing for the current segment calculation.

[0034] The hidden state of the previous segment is expressed as:

[0035] ,

[0036] When calculating the current segment, join the historical memory:

[0037] ,

[0038] 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, Represents the Lth 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 attention calculation; is the key matrix of the previous segment; is the value matrix of the previous section.

[0039] Step 3 includes:

[0040] Step 3-1, scene matching of real-time streaming engine: based on the final user location Query the spatiotemporal clustering results generated by the improved ST-DBSCAN algorithm 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:

[0041] ,

[0042] in is the geographical distance, is the time interval, is the cluster of historical scenes;

[0043] 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:

[0044] ,

[0045] Based on the adjusted interest weight , generate a real-time recommended service list;

[0046] Step 3-3, offline engine deep demand prediction: using the user's long-term behavior pattern library to predict the user's potential needs in the current scenario, and calculate the weight of the user's deep needs through the attention mechanism :

[0047] ,

[0048] in is the similarity score of the user to the requirement γ, is the similarity score of the user to the historical scene δ, and H is the total number of historical scenes;

[0049] Step 3-4, offline engine demand aggregation: Based on the deep demand weight, aggregate the user's historical behavior and tag information to generate a deep demand prediction vector :

[0050] ,

[0051] in A feature representation vector representing the user's historical behavior and tag information in historical scenarios;

[0052] Step 3-5, integration of real-time stream and offline engine: dynamically adjust the weight of real-time stream and 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 recommended services:

[0053] ,

[0054] ,

[0055] in, is the adjustment parameter, For real-time recommendation scores, Score for offline recommendation.

[0056] Step 4 includes:

[0057] Based on the user's current location , query all services within the geo-fence, obtain the access frequency of the service, and calculate the weight of each service , sort by popularity and generate a list of popular services:

[0058] ,

[0059] in, 、 is the weight parameter, 、 are the access frequency and rating of service s, respectively;

[0060] Use the pre-trained meta-policy network to adapt to the behavior pattern of new users. Complete the user state definition based on the user's current geographic location, time, and the list of services the user has interacted with, and define the reward function r:

[0061] ,

[0062] in, Indicates whether the user clicks on the recommended service , It is the user who is serving The dwell time on 、 is the weight parameter;

[0063] Finally, meta-reinforcement learning is used to update the policy network parameters , optimize the recommendation strategy.

[0064] In step 4, the policy network parameters The update formula is:

[0065] ,

[0066] in, 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.

[0067] Step 5 includes: text analysis of meteorological service materials, extracting keywords, and building a meteorological database; defining threshold rules, which are expressed as a set ,in is the xth meteorological element, yes The corresponding threshold;

[0068] When the matching keywords or meteorological elements exceed the threshold, the recommendation mechanism is automatically triggered to generate a list of meteorological services closely related to the user's current environment. :

[0069] ,

[0070] in, is a keyword matching set, is a collection of threshold alarm information;

[0071] By conducting text statistical analysis on meteorological service materials, we build a high-frequency vocabulary in the meteorological database, match current environmental characteristics with keywords in real time, and generate a keyword matching set containing relevant services;

[0072] The meteorological service materials are analyzed using natural language processing technology, including text cleaning, word segmentation, and part-of-speech tagging. Keywords are extracted using the TF-IDF and TextRank algorithms, and semantic expansion is performed using BERT word embedding. A structured meteorological database keyword library containing keywords, categories, weights, semantic vectors, and associated services is constructed. During real-time matching, the spatiotemporal quadruple from step 1 is used to extract current environmental features. The environmental feature labels are compared with the keyword library through exact matching and semantic similarity calculation. Threshold rules are used to prioritize high-weight keywords to generate a keyword matching set. , and mapped to relevant meteorological services to form a real-time, accurate and diverse list of recommended services;

[0073] Based on a set of predefined threshold rules , monitor real-time meteorological data, when meteorological elements Exceeding the threshold When the threshold value is exceeded, the corresponding service is triggered and a threshold alarm information set is generated.

[0074] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0075] The present invention also provides a storage medium storing a computer program or instruction, which executes the steps of the method when the computer program or instruction is run on a computer.

[0076] Beneficial Effects: The method of the present invention integrates data through a dynamic weighting mechanism, dynamically adjusting the weights of the real-time stream and offline engine based on the activity level and time interval of user behavior, thereby generating the final recommendation service. This method fully utilizes multi-dimensional features such as spatiotemporal quadruple data, user tag information, and dynamic interest weights, effectively improving the real-time performance, interpretability, and generalization capabilities of the recommendation system. Furthermore, by introducing multimodal data fusion and diversity constraints, the diversity of recommendation results and user experience are further enhanced. The advantage of this technology lies in its ability to simultaneously meet users' immediate and underlying needs, providing accurate and diverse recommendation services while also possessing strong adaptability and scalability for complex and dynamic scenarios. The advantage of the solution of the present invention lies in providing a fast and accurate initial recommendation experience for new users. The geofencing service ensures the regional relevance of recommendations, while meta-reinforcement learning enhances the personalization of recommendations by rapidly learning user preferences. This combined geofencing and meta-reinforcement learning approach not only solves the cold start problem but also enhances the adaptability and scalability of the recommendation system, making it suitable for dynamic and changing user environments. The present invention can respond to weather changes in real time, providing accurate and timely weather service recommendations, ensuring that users receive reliable information support under dynamic weather conditions. Keyword matching ensures the accuracy of recommended content, and threshold information prompts provide timely warnings, significantly improving user experience and system practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0078] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0079] like Figure 1 As shown, an embodiment of the present invention provides an intelligent service recommendation method based on multi-dimensional scene perception and dynamic portrait modeling, including:

[0080] Step 1: Deploy distributed edge computing nodes to generate spatiotemporal quadruple in real time through GPS / WiFi fingerprint dual-mode positioning and meteorological API ,in Record 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 conditions.

[0081] The GPS and WiFi fingerprint positioning results are fused through Kalman filtering:

[0082] ,

[0083] ,

[0084] in is the final user location, 、 They are the original longitude and latitude coordinates provided by the GPS module and the original longitude and latitude coordinates provided by the WiFi module. is the positioning source variance of the WiFi module, is the positioning source variance of the GPS module, To dynamically calculate the weight coefficient;

[0085] The weather status Mapped into a 6-dimensional vector (temperature / humidity / wind speed / precipitation / PM2.5 / ultraviolet), normalized using a piecewise linear function:

[0086] ,

[0087] in ; Refers to the normalized value of the kth meteorological parameter (the first meteorological parameter is temperature, the second meteorological parameter is humidity, the third meteorological parameter is wind speed, the fourth meteorological parameter is precipitation, the fifth meteorological parameter is PM2.5, and the sixth meteorological parameter is ultraviolet light), refers to the original observation value of the kth meteorological parameter, 、 are the historical mean and standard deviation, respectively.

[0088] The improved Spatio-TemporalDensity-Based Spatial Clustering of Applications with Noise (ST-DBSCAN) algorithm is used to define the spatio-temporal composite distance:

[0089] ,

[0090] in Refers to the i-th point and the jth point The space-time composite distance between Refers to 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;

[0091] if , Is the clustering parameter, then determine the i-th point and the jth point Belong to the same cluster, complete the division of spatiotemporal clustering, and obtain the spatiotemporal clustering results;

[0092] Embed the user tag into the tuple and expand it into a five-tuple form: ,in For the i-th user The user tag of time uses the following formula to calculate the tag weight:

[0093] ,

[0094] in, For users at all times Dynamic weight for category m, is the time interval since the last update, is the weight score of behavior o, (·) is an indicator function used to determine an action Is it part of a collection? , if it is, the value is 1, otherwise the value is 0; A collection of associated behaviors.

[0095] By integrating the above spatiotemporal data and user tag information, the construction of a high-dimensional recommendation database is realized.

[0096] In step 2, a user portrait construction method based on dynamic weight calculation and long-term dependency modeling is proposed. The user interest weight is quantified through a spatiotemporal-aware multi-head attention mechanism, and an improved Transformer-XL architecture is used to capture cross-period behavior patterns.

[0097] The system uses spatiotemporal coding and attention mechanisms to calculate the dynamic interest intensity of users in different service categories, and introduces a time decay factor to achieve accurate modeling of interest evolution. Updated interest weight calculation:

[0098] ,

[0099] in, represents the normalized interest weight of the user on label m at time step t, is the query vector, is the projection matrix of category m, is the scaling factor, is the key matrix, which encodes the semantic information of historical behavior;

[0100] Perform time decay correction calculations:

[0101] ,

[0102] Where, represents the user’s final interest weight 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 decay of the historical interest weight. Based on the final interest weight, the aggregation of user behavior habits and tags is completed;

[0103] 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.

[0104] The hidden state of the previous segment is expressed as:

[0105] ,

[0106] When calculating the current segment, join the historical memory:

[0107] ,

[0108] 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.

[0109] 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:

[0110] 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:

[0111] ,

[0112] in, is the geographical distance, is the time interval, is the cluster of historical scenes.

[0113] 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:

[0114] ,

[0115] Based on the adjusted interest weight , generate a real-time recommended service list;

[0116] Step 3-3, offline engine deep demand prediction: using the user's long-term behavior pattern library to predict the user's potential needs in the current scenario, and calculate the weight of the user's deep needs through the attention mechanism :

[0117] ,

[0118] in is the similarity score of the user to the requirement γ, is the similarity score of the user to the historical scene δ, and H is the total number of historical scenes;

[0119] Step 3-4, offline engine demand aggregation: Based on the deep demand weight, aggregate the user's historical behavior and tag information to generate a deep demand prediction vector :

[0120] ,

[0121] Step 3-5, integration of real-time stream and offline engine: dynamically adjust the weight of real-time stream and 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 recommended services:

[0122] ,

[0123] ,

[0124] in, is the adjustment parameter, Δt is the time interval since the user’s last interaction, For real-time recommendation scores, Score for offline recommendation.

[0125] Step 4: For new users, synchronously call the geofence to obtain regional popular services and implement cold start optimization based on meta-reinforcement learning in parallel; this includes:

[0126] Based on the user's current location , query all services within the geographic fence and obtain information such as service access frequency. , sort by popularity and generate a list of popular services:

[0127] ,

[0128] in, 、 is a weight parameter used to balance the impact of visit frequency and score, 、 are the access frequency and score of service s respectively.

[0129] Based on meta-reinforcement learning, the system will quickly adapt to the behavior patterns of new users and optimize the recommendation strategy. Using the pre-trained meta-strategy network to quickly adapt to the behavior patterns of new users, the user status is defined based on the user's current location, time, and the list of services the user has interacted with. Define the reward function , calculated based on the user's interactive behavior:

[0130] ,

[0131] in, Indicates whether the user clicks on the recommended service ( Indicates the click status of users using the service. The value will be stored in the database). It is the user who is serving The dwell time on 、 is the weight parameter;

[0132] Finally, meta-reinforcement learning is used to update the policy network parameters , optimize the recommendation strategy, the formula is:

[0133] ,

[0134] in, 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, which is used to quantify the performance of the recommendation strategy in the current state.

[0135] Step 5: Establish an intelligent recommendation closed loop associated with weather characteristics: automatically match and trigger relevant weather service recommendations by semantically analyzing keywords and threshold alarm information in weather service materials, including: text analysis of weather service materials, extracting keywords, and building a weather database; at the same time, define threshold rules, which are expressed as a set ,in is the xth meteorological element, yes The corresponding threshold value, the threshold value The value of is based on the statistical analysis of historical meteorological data and expert experience. The specific method is: First, collect the meteorological elements Calculate the mean of historical observation data related to temperature, humidity, wind speed, precipitation, PM2.5, ultraviolet rays, etc. and standard deviation Then, set thresholds based on the business needs of meteorological services and safety warning standards. , where k is the adjustment coefficient, which is usually between 1.5 and 3 and is determined by meteorological experts based on the fluctuation characteristics of specific meteorological elements and application scenarios; for extreme weather scenarios, it is calibrated in combination with the standard thresholds issued by the National Meteorological Administration to ensure Can effectively trigger relevant service recommendations;

[0136] When the matching keywords or meteorological elements exceed the threshold, the recommendation mechanism is automatically triggered to generate a list of meteorological services closely related to the user's current environment. :

[0137] ,

[0138] in, is a keyword matching set, is a collection of threshold alarm information;

[0139] By conducting text statistical analysis on meteorological service materials over the past 10 years, we build a high-frequency vocabulary in the meteorological database, match current environmental characteristics with keywords in real time, and generate a keyword matching set containing relevant services;

[0140] The meteorological service materials are analyzed through natural language processing technology, including text cleaning, word segmentation, and part-of-speech tagging. Keywords are extracted by combining TF-IDF and TextRank algorithms, and semantic expansion is performed by using BERT word embedding. A structured meteorological database keyword library containing keywords, categories, weights, semantic vectors, and related services is constructed, which supports regular updates and dynamic optimization. During real-time matching, the spatiotemporal quadruple of step 1 is used to extract current environmental features. The environmental feature labels are compared with the keyword library through exact matching and semantic similarity calculation. The threshold rules are combined to prioritize triggering high-weight keywords to generate a keyword matching set. , and mapped to relevant meteorological services to form a real-time, accurate and diverse list of recommended services;

[0141] Based on a set of predefined threshold rules , monitor real-time meteorological data, when meteorological elements Exceeding the threshold When the threshold value is exceeded, the corresponding service is triggered and a threshold alarm information set is generated.

[0142] In order to improve the pertinence of meteorological services, this embodiment has developed a personalized recommendation technology route based on user behavior and weather characteristics. User behavior-driven dynamic recommendation technology. Based on user interaction logs, key information such as access time, geographic location, product preferences, and weather query records are extracted to construct a time-space-behavior three-dimensional feature database. On this basis, high-frequency behavior pattern mining and interest weight calculation algorithms are used to achieve dynamic updates of user tags every hour, thereby building a real-time updated user portrait. The system adopts a "real-time stream computing + offline modeling" dual-engine architecture. The real-time module makes instant product recommendations based on the user's current context; 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 uses initial interaction data in the background to quickly build short-term portraits, effectively alleviating the cold start problem.

[0143] The present invention provides an intelligent service recommendation method based on multi-dimensional scene perception and dynamic portrait modeling. There are many methods and approaches to implement this technical solution. The above is only a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. An intelligent service recommendation method based on multi-dimensional scene perception and dynamic portrait modeling, characterized in that: The following steps are involved: Step 1: Deploy distributed edge computing nodes to collect spatiotemporal quadruple data of user operations in real time, store user behaviors and tags, and build a high-dimensional recommendation database. Step 2: Build a dynamic profiling engine, use the attention mechanism to calculate interest weights, and use Transformer to mine long-term associations in behavior sequences to build a library of long-term user behavior patterns. 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. Step 4: For new users, geo-fencing is synchronously called to obtain popular services in the area, and cold start optimization based on meta-reinforcement learning is implemented in parallel; Step 5: Establish a closed loop of intelligent recommendations based on weather characteristics: By semantically analyzing keywords and threshold alarm information in meteorological service materials, relevant meteorological service recommendations are automatically matched and triggered. Step 1 includes: deploying edge computing nodes, generating spatiotemporal quadruple in real time through GPS, WiFi fingerprint dual-mode positioning and meteorological API where t i is the absolute time when the operation of the i-th user occurs, φ i is the i-th user in the WGS-84 coordinate system at t i Longitude value at time, λ i is the i-th user in the WGS-84 coordinate system at t i The latitude value of time, w i Refers to the i-th user at t i Normalized meteorological conditions at the time; Step 1 also includes: fusing the GPS and WiFi fingerprint positioning results through Kalman filtering: in is the final user location, l GPS 、l WiFi They are the original longitude and latitude coordinates provided by the GPS module and the original longitude and latitude coordinates provided by the WiFi module. is the positioning source variance of the WiFi module, is the positioning source variance of the GPS module, and α is the dynamic calculation weight coefficient; The weather status w i Mapped into a 6-dimensional vector: temperature, humidity, wind speed, precipitation, PM2.5, and ultraviolet light, and normalized using a piecewise linear function: where k = 1,...,6; Refers to the normalized value of the kth meteorological parameter, w (k) Refers to the original observation value of the kth meteorological parameter, μ k , σ k are the historical mean and standard deviation respectively; The improved spatial clustering ST-DBSCAN algorithm based on noise application of spatiotemporal density is used to define the spatiotemporal composite distance: Among them, D ST (p Δi ,p Δj ) refers to the Δi-th point p Δi and the Δj-th point p Δj The space-time composite distance between geo (·) refers to geographical distance, T max is the time normalization coefficient, β is the spatial weight factor, Respectively The time scalar and A time scalar; If D ST (p Δi ,p Δj )≤∈,∈ is the clustering parameter, then the Δi-th point p Δi and the Δj-th point p Δj Belong to the same cluster, complete the division of spatiotemporal clustering, and obtain the spatiotemporal clustering results; Embed the user tag into the tuple and expand it into a five-tuple form: <t i ,φ i ,λ i ,w i ,τ i >, where τ i For the i-th user at time t i The user tag is calculated using the following formula: in, For the user at time t i The dynamic weight of label m, Δt is the time interval from the last update, action o Represents user behavior o, δ o is the weight score of behavior o, I(·) is the indicator function, is a set of associated behaviors, e is a natural constant; Finally, the spatiotemporal quadruple and user tag information are integrated to build a high-dimensional recommendation database; Step 2 includes: using spatiotemporal coding and attention mechanism to calculate the dynamic interest intensity of users in different service categories, introducing a time decay factor, and the updated interest weight calculation formula is: in, represents the normalized interest weight of the user on label m at time step t, is the query vector, W (m) is the projection matrix of label m, is the scaling factor, K is the key matrix; Perform time decay correction calculations: in, represents the user's final interest weight for tag m at time t, and λ is the decay rate; Based on the final interest weight, complete the aggregation of user behavior habits and tags; The five-tuple from step 1 is fed into the Transformer-XL, a very long context transformer. The Transformer-XL includes a recurrent memory module and relative position encoding to identify long-term user behavior patterns. The recurrent memory module is responsible for caching the hidden state of the previous segment and completing historical memory splicing for the current segment calculation. The hidden state of the previous segment is expressed as: M τ-1 =[h τ-1,1 ,...,h τ-1,L ], When calculating the current segment, join the historical memory: Among them, M τ-1 It is the memory cache of the previous section, which is composed of the hidden state h of each layer in the previous section. τ-1,1 Composition, L is the sequence length, h τ-1,L Represents the Lth hidden state in the previous segment; is the attention output of the current segment, Q τ , K τ 、V τ are the query, key, and value matrices of the current segment respectively; Attention represents attention calculation; K τ-1 is the bond matrix of the previous segment; V τ-1 is the value matrix of the previous section; Step 3 includes: Step 3-1, scene matching of real-time streaming engine: based on the final user location Query the spatiotemporal clustering results generated by the improved spatiotemporal density-based noise application spatial clustering ST-DBSCAN algorithm in step 1 to find the cluster C to which the user's current scene belongs. current ; Calculate the similarity Sim(C current ,C history ), the formula is: where d spatial is the geographical distance, d tcmporal is the time interval, C history is the cluster of historical scenes; 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 real-time recommended service list; Step 3-3, offline engine deep demand prediction: Utilize the user's long-term behavior pattern library to predict the user's potential needs in the current scenario, and calculate the weight γ of the i-th user's deep demand through the attention mechanism i : where s γ is the similarity score of the user to the requirement γ, s δ is the similarity score of the user to the historical scene δ, and H is the total number of historical scenes; Step 3-4, offline engine demand aggregation: According to the deep demand weight, aggregate the user's historical behavior and tag information to generate the deep demand prediction vector v deep ; Step 3-5, integration of real-time stream and offline engine: dynamically adjust the weight α of real-time stream and offline engine according to the activity level and time interval of user behavior adjusted , and then according to the final recommendation score S final , generate a list of recommended services: S final =a adjusted ·S real-time +(1-a adjusted )·S offlinc , Among them, Δβ is the adjustment parameter, S real-time is the real-time recommendation score, S offlinc Score for offline recommendations; Step 4 includes: Based on the user's current location Query all services within the geofence, obtain the access frequency of the services, and calculate the weight w of each service. s , sort by popularity and generate a list of popular services: Among them, θ, is a weight parameter, popularity(s) and rating(s) are the access frequency and rating of service s, respectively; Use the pre-trained meta-policy network to adapt to the behavior pattern of new users. Complete the user state definition based on the user's current geographic location, time, and the list of services the user has interacted with, and define the reward function r: r=ρ·click(a)+σ·duration(a), Among them, click(a) indicates whether the user clicks on the recommended service a, duration(a) is the time the user stays on service a, and ρ and σ are weight parameters; Finally, meta-reinforcement learning is used to update the policy network parameters θ new , optimize the recommendation strategy; In step 4, the policy network parameters θ new The update formula is: Among them, θ meta are the pre-trained meta-policy network parameters, η is the learning rate, is the gradient of the reward, r t represents the value of the reward function at time step t; Step 5 includes: performing text analysis on meteorological service materials, extracting keywords, and building a meteorological database; defining threshold rules, which are represented by a set R = {(e x ,v x )}, where e x is the xth meteorological element, v x It is e x The corresponding threshold; When the matching keywords or meteorological elements exceed the threshold, the recommendation mechanism is automatically triggered to generate a list of meteorological services closely related to the user's current environment S recommend : Among them, M keyword is the keyword matching set, M threshold is a collection of threshold alarm information; By conducting text statistical analysis on meteorological service materials, we build a high-frequency vocabulary in the meteorological database, match current environmental characteristics with keywords in real time, and generate a keyword matching set containing relevant services; The meteorological service materials are analyzed through natural language processing technology, including text cleaning, word segmentation, and part-of-speech tagging. Keywords are extracted by combining TF-IDF and TextRank algorithms, and semantic expansion is performed by using BERT word embedding to build a structured meteorological database keyword library containing keywords, categories, weights, semantic vectors, and associated services. During real-time matching, the spatiotemporal quadruple of step 1 is used to extract current environmental features. The environmental feature labels are compared with the keyword library through exact matching and semantic similarity calculation. The threshold rule is combined to give priority to triggering high-weight keywords to generate a keyword matching set M. keyword , and mapped to relevant meteorological services to form a real-time, accurate and diverse list of recommended services; Based on the predefined threshold rule set {(e x ,v x )}, monitor real-time meteorological data, when meteorological elements e x Exceeding the threshold v x When the threshold value is exceeded, the corresponding service is triggered and a threshold alarm information set is generated.

2. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor is caused to perform the steps of the method according to claim 1 .

3. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to claim 1 are executed.

Images