Short message template intelligent matching and effect prediction optimization method for industry scene

Abstract

The invention relates to the technical field of short message sending, in particular to an industry scene-oriented short message template intelligent matching and effect prediction optimization method, which is applied to a short message service platform carrying a machine learning inference engine. Industry scene features and user response attenuation features are extracted through block processing and a community detection algorithm; and based on scene time sequence evolution analysis and in combination with effect stage modeling, dynamic prediction of short message effect attenuation is realized. And then constructing an industry knowledge graph, predicting the potential effect risk of the template, performing semantic association matching in combination with a multi-strategy library, generating an effect risk management plan, and finally realizing real-time dynamic optimization of the short message template. According to the method, the scene adaptability, the prediction accuracy and the strategy matching degree are effectively improved, and the short message marketing effect and the resource utilization rate are remarkably optimized.

Description

Technical Field

[0001] This invention relates to the field of SMS sending technology, and in particular to an intelligent matching and effect prediction optimization method for SMS templates for industry scenarios. Background Technology

[0002] In the field of digital marketing, the core technical challenge of intelligent matching and performance optimization of SMS templates lies in the fundamental contradiction between the dynamic evolution of industry scenarios and the static adaptation mechanism of SMS templates. Existing SMS service platforms generally rely on fixed rules or historical experience for template matching, lacking the ability to dynamically capture the temporal evolution of industry scenarios, leading to the following technical bottlenecks: On the one hand, data collection and preprocessing failed to design adaptive strategies for the dynamic nature of scenarios, making it difficult to accurately extract industry scenario characteristics and user response decay characteristics. Industry scenario characteristics include semantic tags for scenarios such as e-commerce promotions and financial risk control notifications, while user response decay characteristics include time-sensitivity and content fatigue. On the other hand, the effect prediction model did not deeply integrate the coupling relationship between scenario time-series data and user behavior data, failing to effectively analyze the fluctuation intensity, evolution trajectory, and time-series periodicity of scenario feature points. This resulted in predictions of SMS effect decay remaining at a rough statistical level, making it difficult to predict the potential risks of different templates in dynamic scenarios. Furthermore, the matching of optimization strategies with scenario requirements relied on manual experience mapping, lacking semantic association analysis based on knowledge graphs. This prevented strategies such as template content adjustment and sending timing selection from forming intelligent linkages with real-time scenario characteristics. These problems directly lead to a disconnect between SMS content and current user needs, causing industry pain points such as decreased response rates and wasted marketing resources. There is an urgent need for a technical solution that can dynamically perceive scenario changes, intelligently predict effect risks, and achieve adaptive optimization of template strategies.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide a method for intelligent matching and performance prediction optimization of SMS templates for industry scenarios. This method aims to solve the fundamental contradiction between the dynamic evolution of industry scenarios and the static adaptation mechanism of SMS templates in the field of digital marketing. This contradiction leads to the inability of data collection, performance prediction, and optimization strategy matching to dynamically respond to changes in scenarios, resulting in a disconnect between SMS content and user needs, a decline in response rate, and a waste of marketing resources.

[0005] To achieve the above objectives, this invention provides a method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios, applied to an SMS service platform, wherein the SMS service platform is equipped with a machine learning inference engine; the method includes: The machine learning inference engine of the SMS service platform uses a data collection strategy to obtain SMS sending records and user interaction behavior data, and preprocesses the data to obtain preprocessed SMS data. The preprocessed SMS data is divided into blocks to obtain the block-processed SMS data; based on the community detection algorithm, the block-processed SMS data is classified according to industry scenario characteristics and user response decay characteristics to obtain target industry scenario data and target user response decay data. Based on the target industry scenario data, scenario time-series evolution analysis is performed to obtain scenario time-series evolution data; based on the scenario time-series evolution data and target user response decay data, effect stage modeling is used to predict SMS effect decay and obtain effect decay prediction data. Construct an industry knowledge graph, and use the industry knowledge graph to predict the potential effect risk of SMS templates based on the scenario time-series evolution data and effect decay prediction data to obtain potential effect risk data. Several SMS optimization strategy libraries are obtained from the database. Semantic association matching is performed on the potential effect risk data and each SMS optimization strategy to obtain semantic association matching results. An effect risk management plan is generated based on the semantic association matching results. The SMS template is dynamically optimized based on the effect risk management plan.

[0006] Furthermore, to achieve the above objectives, the present invention also provides an intelligent matching and effect prediction optimization device for SMS templates in industry scenarios. The device includes: a memory, a processor, and an intelligent matching and effect prediction optimization program for SMS templates in industry scenarios stored on the memory and executable on the processor. The intelligent matching and effect prediction optimization program for SMS templates in industry scenarios is configured to implement the steps of the intelligent matching and effect prediction optimization method for SMS templates in industry scenarios as described above.

[0007] In addition, to achieve the above objectives, the present invention also provides a medium storing an industry-specific intelligent matching and effect prediction optimization program for SMS templates, wherein when the industry-specific intelligent matching and effect prediction optimization program is executed by a processor, it implements the steps of the industry-specific intelligent matching and effect prediction optimization method for SMS templates as described above.

[0008] This invention provides an intelligent matching and effect prediction optimization method for SMS templates in industry scenarios. The method utilizes an adaptive data acquisition strategy and community detection algorithm of a machine learning inference engine to accurately extract industry scenario features and temporal evolution patterns, overcoming the limitations of traditional static rule matching. This achieves dynamic semantic alignment between SMS templates and real-time scenarios, improving the timeliness and accuracy of scenario feature capture. Based on the coupled analysis of scenario temporal evolution data and user response decay characteristics, a dynamic decay prediction model is constructed using effect stage modeling. This model quantifies key factors such as scenario fluctuation intensity and user fatigue, solving the problem of lagging prediction of effect decay trends in traditional statistical models. This enables the prediction of SMS effects from high response levels... Accurate full-cycle prediction from the initial to the decay period; by mining deep semantic relationships between scene features, effect prediction data, and template risks through industry knowledge graphs, overcoming the limitations of manual experience mapping, identifying potential risks of templates in specific scenarios in advance, and generating personalized optimization strategies based on semantic relationship matching, forming a closed-loop management mechanism of risk prediction-strategy generation-dynamic optimization; through the dynamic optimization mechanism, intelligent linkage between SMS templates and real-time user needs is achieved, significantly reducing user response fatigue caused by scenario mismatch, improving click-through rate and conversion rate, while reducing ineffective strategy investment, making the allocation of marketing resources more in line with the dynamic needs of the scenario, and promoting the transformation of digital marketing from experience-driven to data intelligence-driven. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an embodiment of the intelligent matching and effect prediction optimization method for SMS templates for industry scenarios according to the present invention.

[0010] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0011] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0012] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the intelligent matching and effect prediction optimization method for SMS templates in industry scenarios according to the present invention. An embodiment of the intelligent matching and effect prediction optimization method for SMS templates in industry scenarios according to the present invention is presented.

[0013] In one embodiment, the intelligent matching and effect prediction optimization method for SMS templates for industry scenarios is applied to an SMS service platform, which is equipped with a machine learning inference engine; the method includes: In step S100, the machine learning inference engine of the SMS service platform uses a data collection strategy to obtain SMS sending records and user interaction behavior data, and preprocesses the data to obtain preprocessed SMS data.

[0014] The SMS service platform can be a digital marketing infrastructure system providing SMS sending, template management, and performance monitoring capabilities, supporting the entire process of SMS template matching, sending, and performance optimization. The machine learning inference engine can be an intelligent computing module deployed within the SMS service platform, used to perform model inference and strategy decision-making. Its operating principle involves loading pre-trained models and real-time data streams for online inference, and collaborating with data acquisition strategies to dynamically capture raw data. The data acquisition strategy can be a set of data acquisition rules that adaptively adjusts based on the dynamics of the scenario. Its operating principle is to dynamically adjust the collection frequency, field range, and data source weight based on historical scenario evolution trends and user response patterns. SMS sending records can be log data containing metadata such as SMS template content, sending time, target user group, and industry tags, used as basic input data for scenario feature extraction and performance analysis. User interaction behavior data can be time-series data of user feedback behaviors such as clicking, ignoring, unsubscribing, and converting SMS messages, used to characterize user response decay features. Preprocessed SMS data can be data that has undergone cleaning, noise reduction, standardization, and structuring, the technical effect of which is to improve the accuracy and efficiency of subsequent block processing and feature classification. Data preprocessing can involve performing operations such as missing value imputation, outlier filtering, field alignment, and format standardization. Furthermore, rule-based cleaning processes combined with statistical distribution verification or the use of graph neural networks can be used to complete the context of user behavior sequences, thereby eliminating noise interference and unifying the data structure.

[0015] Step S200: The preprocessed SMS data is divided into blocks to obtain block-processed SMS data. Based on the community detection algorithm, the block-processed SMS data is classified according to industry scenario characteristics and user response decay characteristics to obtain target industry scenario data and target user response decay data.

[0016] The segmented SMS data can be preprocessed subsets divided by time window, industry category, or user group. This facilitates community detection algorithms in identifying high-cohesion features within a local semantic space. Segmenting the preprocessed SMS data can involve dividing the data into mutually exclusive subsets according to preset dimensions. Further, it can be done through continuous segmentation based on a time sliding window or discrete segmentation based on industry labels, thereby reducing the computational complexity of the community detection algorithm. The community detection algorithm can be a graph theory algorithm used to identify highly cohesive subgraphs in complex networks. Its operating principle is to construct the SMS data as a node-edge relationship graph, identify semantically consistent subgroups through modularity optimization or spectral clustering methods, and collaboratively build a decay prediction model with the effect stage modeling. Industry scenario features can be a set of semantic labels representing the business status of a specific industry at a specific time period, including but not limited to promotional activity features, compliance notification features, and service reminder features. User response decay features can be behavioral indicators reflecting the decline in users' willingness to respond to similar SMS messages over time or frequency, including but not limited to content fatigue, time-period sensitivity, and frequency tolerance threshold. Data classification of segmented SMS data based on community detection algorithms can be achieved by constructing a heterogeneous graph from the segmented data, running the community detection algorithm to identify highly cohesive subgraphs and mapping them to feature labels. Furthermore, feature classification can be performed by constructing a user-template-scenario tripartite graph and applying multi-layer community detection, or by using density clustering as an alternative graph algorithm in the embedding space. This effectively separates semantically consistent industry scenarios from user decay patterns. Target industry scenario data can be a subset of data representing semantic consistency within a specific industry scenario after classification by the community detection algorithm, serving as the input basis for scenario temporal evolution analysis. Target user response decay data can be a subset of data representing response decay patterns for a specific user group after classification by the community detection algorithm, used to couple with scenario temporal evolution data to support effect decay prediction modeling.

[0017] Step S300: Perform scenario time-series evolution analysis based on target industry scenario data to obtain scenario time-series evolution data. Based on the scenario time-series evolution data and target user response decay data, use effect stage modeling to predict SMS effect decay and obtain effect decay prediction data.

[0018] Scene temporal evolution analysis can be a process of time-series modeling of target industry scene data to identify its evolutionary patterns. Its technical effect is to extract dynamic attributes such as scene fluctuation intensity, periodicity, and temporal trajectory. Scene temporal evolution data can be structured data containing the change trajectory, fluctuation intensity, and periodic patterns of scene feature points over time, serving as a key input for effect stage modeling. Scene temporal evolution analysis based on target industry scene data can involve sorting the data by time and applying temporal decomposition or pattern mining algorithms to extract evolutionary patterns. Furthermore, it can quantify scene fluctuation intensity and periodicity by using STL decomposition to extract trends and seasonal terms or by employing a Transformer encoder to model long-term dependencies. Effect stage modeling can be a predictive framework that divides the SMS lifecycle into high-response, stable, and decay stages and models them separately. Its operating principle is to integrate scene temporal evolution data and user response decay data, using temporal neural networks or state transition models for stage identification and effect prediction, including but not limited to Hidden Markov Stage Models, LSTM Stage Regression Models, and Change Point Detection Stage Division Models. Modeling effect stages based on scenario time-series evolution data and target user response decay data involves using these two types of data as input features to train a stage identification model and outputting predicted effect values ​​for each stage. Furthermore, survival analysis models can be used to estimate user response duration, or a multi-task learning framework can be constructed to simultaneously predict stage attribution and response rate, thus achieving a leap from extensive statistics to full-cycle dynamic decay prediction. Effect decay prediction data can be a set of prediction results including predicted response rates, decay rates, inflection point times, and other indicators for each stage. The technical effect is that this data is input into an industry knowledge graph for risk association analysis.

[0019] Step S400: Construct an industry knowledge graph. Based on scenario time-series evolution data and effect decay prediction data, use the industry knowledge graph to predict the potential effect risk of SMS templates and obtain potential effect risk data.

[0020] Industry knowledge graphs can be knowledge bases organized in the form of triples, encompassing semantic relationships between industry scenarios, template content, performance metrics, and risk factors. Their operational principle involves constructing knowledge graphs from historical data and expert rules through entity extraction, relation mining, and graph embedding techniques. Examples include, but are not limited to, knowledge graphs for e-commerce, financial risk control, and logistics services. Building an industry knowledge graph involves extracting entities and relations from historical templates, scenario logs, and expert rules, constructing a triple graph, and embedding it. Further, it can be achieved by extracting relations using BERT+OpenIE and constructing an RDF graph, or by using graph neural networks to learn node embeddings and relation weights end-to-end, thereby establishing a reasonable semantic network between scenarios, performance, and risk. Predicting potential performance risks based on scenario time-series evolution data and performance decay prediction data involves using the predicted data as query conditions to perform path reasoning or similar node retrieval in the knowledge graph. Further, it can be achieved by using graph attention mechanisms to weighted aggregate neighbor risk signals or by calculating the semantic mismatch degree of template-scenario pairs through knowledge graph embedding, thereby identifying potential failure risks of templates in specific scenarios in advance. Potential effect risk data can be risk assessment results derived from knowledge graph reasoning, which can be used to drive semantic matching and contingency plan generation for optimization strategies.

[0021] Step S500: Obtain several SMS optimization strategy libraries from the database, perform semantic association matching on potential effect risk data and each SMS optimization strategy, obtain semantic association matching results, generate effect risk management plan based on semantic association matching results, and dynamically optimize SMS templates based on effect risk management plan.

[0022] The SMS optimization strategy library can be a collection of strategies storing various executable optimization actions and their applicable conditions, including but not limited to template content rewriting strategies, sending timing adjustment strategies, and user segmentation and redirection strategies. Retrieving the SMS optimization strategy library from the database can involve loading a predefined set of optimization strategies. Furthermore, it can be done by loading industry-specific strategy subsets or filtering relevant strategy clusters based on risk type, thus providing structured candidate solutions for semantic matching. Semantic association matching of potential effect risk data and SMS optimization strategies can involve encoding risk descriptions and strategy descriptions into vectors, calculating cosine similarity, or using a cross-attention mechanism for scoring. Further, it can be done by using Sentence-BERT for semantic vectorization and matching, or by constructing a dual-tower model to jointly learn the risk-strategy matching function, thereby automatically linking the most relevant optimization strategies. Generating effect risk management plans based on semantic association matching results can involve selecting strategies with matching scores higher than a threshold, generating executable plans according to priority or combination rules. Further, it can be done by using reinforcement learning to select the optimal strategy combination or by merging complementary strategies based on a rule engine to generate composite plans, thus forming a structured and implementable risk response plan. Dynamic optimization based on the effect risk management plan can involve calling the template editing interface to adjust content variables, sending time, or target user groups. Furthermore, it can be achieved by verifying the optimized version through A / B testing before full deployment or by replacing template variables in real time and immediately triggering the sending process, thereby realizing intelligent linkage between SMS templates and users' real-time needs.

[0023] Taking the optimization of promotional SMS messages during major e-commerce platform promotions as an example, the intelligent matching and effect prediction optimization method for SMS templates for industry scenarios in this embodiment can be as follows: Before the promotion, the machine learning inference engine captures a large number of promotional SMS sending records and user click behaviors through an adaptive data collection strategy; after preprocessing and segmentation, the community detection algorithm identifies the characteristics of limited-time discount scenarios and the attenuation characteristics of users with high-frequency reception fatigue; the scenario time-series evolution analysis finds that the scenario exhibits a strong 72-hour cycle fluctuation; the effect stage modeling predicts that the current template will enter a rapid attenuation period after 24 hours; the industry knowledge graph associates the scenario with the risk of content homogenization and matches two optimization strategies: adding scarcity language and delaying the sending by 2 hours; the system automatically generates a risk management plan and dynamically adjusts the template content and sending timing to maintain a high click-through rate the next day and avoid resource waste.

[0024] This embodiment provides an intelligent matching and performance prediction optimization method for SMS templates in industry scenarios. It accurately captures raw data by executing an adaptive data collection strategy through a machine learning inference engine, separates industry scenario features from user response decay features using a community detection algorithm, and achieves full-cycle dynamic decay prediction by coupling scenario temporal evolution analysis and performance stage modeling. It constructs an industry knowledge graph to mine deep semantic associations, and finally generates optimization plans through semantic association matching and drives dynamic adjustments. This method can significantly improve the alignment accuracy between templates and users' real-time needs, reduce response fatigue, increase click-through rate and conversion rate, and reduce ineffective marketing investment, achieving a high degree of synergy between resource allocation and dynamic scenario needs.

[0025] In one embodiment, a data acquisition strategy is used to acquire SMS sending records and user interaction behavior data, and the data is preprocessed to obtain preprocessed SMS data, including: The data collection cycle and data collection granularity are determined based on preset industry labels.

[0026] The preset industry tags can be predefined semantic classification tags used to identify the business domain to which the SMS belongs, and can be used as contextual basis for driving adaptive adjustments to the data collection strategy. Furthermore, the preset industry tags can include, but are not limited to, one or more of e-commerce promotion tags, financial risk control tags, and logistics notification tags. The data collection cycle can be the time interval between two consecutive data collection operations, and can be used to control the frequency of data capture. In an exemplary embodiment, the data collection cycle can be described in context, i.e., it is dynamically coupled with the data collection granularity to jointly determine the data collection threshold. The data collection granularity can be the fine-grainedness of the data dimensions covered in a single collection or the level of detail in the records, and can be used to determine the richness and structural depth of the collected content. Furthermore, the data collection granularity can be described in context regarding the acquisition method, such as whether it includes extended fields such as user device information and geographical location.

[0027] Determining the data collection cycle and granularity based on preset industry tags can be achieved by querying the configuration rules corresponding to the preset industry tags and dynamically setting the collection time interval and data detail level. Furthermore, this operation can be implemented by indexing a pre-stored cycle-granularity configuration table through industry tags, or by calling a reinforcement learning agent based on industry tags to recommend optimal collection parameters in real time. This allows the data collection parameters to match the inherent rhythm of the industry scenario, improving data relevance.

[0028] A data acquisition threshold is constructed based on the dynamic coupling relationship between the data acquisition cycle and the data acquisition granularity.

[0029] The dynamic coupling relationship can be a non-linear association rule that adaptively adjusts the data acquisition cycle and data acquisition granularity according to changes in industry scenarios. This can be used to ensure simultaneous improvement in acquisition frequency and data detail in highly volatile scenarios. Furthermore, the dynamic coupling relationship can be described in context, i.e., a mapping function trained on historical scenario response data, or conditional logic defined by expert rules. The data acquisition threshold can be a criterion indicator derived from the dynamic coupling relationship between the data acquisition cycle and granularity, triggering high-quality data capture. It can be used as a key input for generating data acquisition strategies. Furthermore, the data acquisition threshold can be described in context, for example, calculated through a coupling function, to determine whether to activate a high-density acquisition mode.

[0030] Constructing a data acquisition threshold based on the dynamic coupling relationship between the data acquisition cycle and the data acquisition granularity can be achieved by substituting the current cycle and granularity into a predefined coupling function and outputting a threshold value. Furthermore, this operation can be implemented by constructing the threshold using a linear weighted combination, or by fitting historical effective acquisition points through a neural network to output a dynamic threshold, thereby forming a quantifiable basis for determining whether to increase the acquisition intensity.

[0031] The data acquisition strategy is determined based on the data acquisition threshold.

[0032] Among them, the data acquisition threshold can be a criterion for triggering high-quality data capture and can be used to determine whether to activate the high-density acquisition mode.

[0033] Determining a data acquisition strategy based on data acquisition thresholds can involve activating a high-frequency, high-granularity acquisition mode when real-time scene metrics exceed the threshold. Furthermore, this operation can be achieved by triggering full-field acquisition and shortening the acquisition interval to minutes when the threshold exceeds the limit, or by using a sliding window to compare thresholds and dynamically adjust the allocation ratio of acquisition resources. This allows for a shift from static acquisition to scene-aware adaptive acquisition.

[0034] The SMS data is subjected to noise filtering to obtain the noise-filtered SMS data.

[0035] The SMS data can be a collection of raw SMS sending records and user interaction behavior data acquired through a data collection strategy but not yet cleaned, which can be used as input for noise filtering. Noise filtering can be a cleaning process that identifies and removes invalid, abnormal, or interfering records from the raw SMS data, and can be used to remove noise such as accidental touches, bot clicks, and duplicate reports. Furthermore, noise filtering can include, but is not limited to, one or more of the following: behavior anomaly detection filtering, timestamp validity verification, and user identity validity verification.

[0036] Noise filtering of SMS data can be achieved by applying rule engines or anomaly detection models to identify and remove data records that do not conform to normal interaction patterns. Furthermore, this operation can be implemented by detecting and filtering abnormal user behavior based on the Isolation Forest algorithm, or by using session duration and click location rationality rules to filter out accidental touches, thereby improving data quality and model stability in subsequent processing stages.

[0037] Feature alignment is performed on the noise-filtered SMS data to obtain preprocessed SMS data.

[0038] The SMS data after noise filtering can be an intermediate data form with invalid records removed, which can be used as input for feature alignment processing. Feature alignment processing can be an operation to standardize and unify multi-source heterogeneous SMS data in terms of time, semantics, and metrics, and can be used to achieve structural consistency across modalities. Furthermore, feature alignment processing can include, but is not limited to, one or more of the following: time alignment, semantic encoding alignment, and metric normalization alignment.

[0039] Feature alignment of noise-filtered SMS data can unify the time base, field naming, and response measurement units of data from different sources. Furthermore, this operation can be achieved by aligning all event logs using a unified timestamp, or by mapping click behaviors from different channels to the same semantic ID and normalizing response latency metrics. This can eliminate feature bias caused by multi-source heterogeneity and ensure consistency in subsequent community detection and modeling.

[0040] Taking data collection optimization in the context of anti-fraud notifications in the financial industry as an example, the intelligent matching and effect prediction optimization method for SMS templates in this embodiment can be as follows: The system identifies that the current SMS belongs to the financial risk control preset industry label, automatically sets a short data collection period (5 minutes) and high granularity (including device fingerprint, IP address, and operation path); the two generate a high-sensitivity data collection threshold through dynamic coupling; when it is detected that the unsubscription rate of a certain template suddenly increases by 200% and exceeds the threshold, it immediately triggers full field collection and shortens the interval to the minute level; after the original SMS data is filtered to remove test accounts and crawler traffic, the clicks in the bank APP and the clicks on the SMS links are uniformly encoded into valid conversions through feature alignment processing, and finally high-quality preprocessed data is output to support subsequent accurate risk prediction.

[0041] This embodiment determines the data acquisition cycle and granularity based on preset industry labels, constructs a data acquisition threshold based on dynamic coupling relationships, generates an adaptive data acquisition strategy based on the threshold, removes interference signals through noise filtering, and finally achieves structured alignment through feature alignment. This fundamentally solves the problem of data acquisition and preprocessing failing to design adaptive strategies for scene dynamics. Specifically, this method provides a high signal-to-noise ratio, semantically consistent, and timely data foundation for subsequent modeling through industry label-driven adaptive parameter configuration, dynamic threshold-triggered acquisition intensity adjustment, multi-dimensional noise filtering, and cross-source feature standardization, significantly improving scene perception sensitivity and template matching accuracy.

[0042] In one embodiment, the preprocessed SMS data is divided into blocks to obtain the block-processed SMS data, including: setting a dynamic time window, and using the continuity of scene boundaries based on the dynamic time window to divide the preprocessed SMS data into blocks to obtain the block-processed SMS data.

[0043] The dynamic time window can be a time segmentation mechanism that adaptively adjusts the window length based on scene activity, event density, or semantic mutation points, enabling data segmentation to align with the actual evolution rhythm of industry scenarios. Furthermore, the dynamic time window can dynamically determine start and end times by monitoring changes in industry tag frequency, user response fluctuations, or semantic embedding distance in preprocessed SMS data. In this embodiment, the dynamic time window can include, but is not limited to, one or more of the following: a sliding window based on event density, an adaptive window based on semantic mutation detection, and an elastic window based on response variance thresholds.

[0044] Scene boundary continuity can be used as a criterion to identify semantically consistent and stable user response patterns as valid scene units, ensuring that each block has highly homogeneous scene semantics and user response features. Furthermore, scene boundary continuity can be combined with calculating the similarity of scene feature vectors between adjacent time periods and the KL divergence of user behavior distribution to determine whether they belong to the same semantic scene. In a specific embodiment, scene boundary continuity may include, but is not limited to, semantic consistency boundaries, behavior stability boundaries, and event coherence boundaries.

[0045] Setting a dynamic time window can be based on scene activity indicators or semantic change signals in preprocessed SMS data, calculating and setting the time span of the current block in real time. Furthermore, setting a dynamic time window can trigger a window reset based on the rate of change of industry label entropy within the sliding window, or use LSTM to predict the semantic drift probability at the next moment to dynamically expand or truncate the window, thereby achieving the technical effect of matching the time window length with the actual evolution speed of the industry scene.

[0046] Based on dynamic time windows, segmenting preprocessed SMS data using scene boundary continuity can be achieved by merging or splitting time periods within a dynamic time window, combined with scene boundary continuity criteria, to form semantically coherent data blocks. Furthermore, this operation can be further optimized by dividing the time series into candidate segments and using graph segmentation algorithms to maximize intra-segment similarity, or by employing online clustering methods to terminate the current block and start a new block when the semantic distance between new data and the current block exceeds a threshold. This generates data blocks with consistent internal semantics and clear external boundaries, providing high-quality input for community detection algorithms. Taking the dynamic segmentation of the end-of-month bill notification cycle in the financial industry as an example, the intelligent matching and effect prediction optimization method for SMS templates in this embodiment can be as follows: During the last 7 days of the month, bill notification SMS messages are sent intensively, but the user response pattern differs significantly between the "billing date - repayment reminder date" and the "overdue collection date". The system, through a dynamic time window mechanism, detected a sudden increase in user unsubscription rate and a sharp drop in click-through rate on the 4th day. Combined with the continuity criterion for scene boundaries, the system divided the first 3 days into a "normal notification period" block and the following 4 days into a "collection-sensitive period" block. This segmentation allows the community detection algorithm to extract features from both types of scenes and their corresponding user fatigue patterns, supporting subsequent accurate prediction of effect decay and template optimization.

[0047] This embodiment sets a dynamic time window to adjust the time span in real time based on scene activity indicators or semantic change signals. By using scene boundary continuity criteria to merge or split time periods based on the dynamic time window, and combining feature vector similarity and KL divergence of user behavior distribution to determine the boundary, it can avoid semantic fragmentation or redundancy caused by a fixed window, preserve the inherent structure of industry scene evolution, provide high-quality input for subsequent community detection algorithms, and thus improve the technical effect of improving the accuracy and granularity of scene temporal evolution analysis.

[0048] In one embodiment, a community detection algorithm is used to classify the segmented SMS data based on industry scenario features and user response decay features to obtain target industry scenario data and target user response decay data. This includes: semantically labeling the segmented SMS data with feature nodes to obtain several semantic labeling categories; obtaining the feature association matrix of each semantic labeling category; constructing a community network model based on the feature association matrix using interaction strength weight quantization; dividing the segmented SMS data into communities based on the community network model to obtain community division results; and using a hybrid classification method based on industry scenario topology and user response decay paths to classify the data based on industry scenario features and user response decay features using the community division results to obtain target industry scenario data and target user response decay data.

[0049] In this context, semantic annotation of feature nodes can be a process of assigning semantic labels with business meaning to each data unit in the segmented SMS data, mapping the raw data to an interpretable semantic space. Furthermore, semantic annotation of feature nodes can include, but is not limited to, one or more of industry intent labels, user status labels, and content topic labels. Semantic label categories can be discretized, mutually exclusive, or overlapping sets of semantic labels formed after semantic annotation of feature nodes, serving as basic units for constructing feature association matrices. It is understandable that the formation process of semantic label categories, as basic units for constructing feature association matrices, can be explained in context by statistically analyzing the joint occurrence of different semantic categories in the same time window, the same user, or the same batch in historical data.

[0050] The feature association matrix can be a two-dimensional matrix with semantic label categories as row and column indices and element values ​​representing the co-occurrence frequency or interaction strength between categories. It can be used to quantify the coupling relationship between different semantic categories. Furthermore, the element values ​​of the feature association matrix can be explained in context by statistically analyzing and normalizing the number of times different semantic categories co-occur in the same context. Interaction strength weights can be normalized numerical indicators used to measure the tightness of interaction between different semantic label categories, and can be used to assign semantic importance to edges in community network models. It is understood that the interaction strength weights, in context, are obtained based on the original frequencies in the feature association matrix, weighted by time decay factors, user group size, or response differences.

[0051] Community network models can be weighted graph structures with semantically labeled categories as nodes and interaction strength weights as edges. They can be used to explicitly model high-order semantic relationships between industry scenarios and user behavior patterns. Furthermore, the acquisition method of a community network model can be explained in context: it is constructed by quantizing the feature association matrix through interaction strength weights, where nodes represent semantic categories and edges represent their dynamic coupling strength. It is understandable that, as the input graph structure for community detection algorithms, the topological quality of the community network model directly affects the semantic consistency of the community segmentation results.

[0052] Semantic annotation of feature nodes is performed on the segmented SMS data to obtain several semantic annotation categories. This can be achieved by assigning one or more semantic labels to each SMS record or user behavior sequence through a rule engine or lightweight classification model. Furthermore, this operation can be implemented by using predefined keyword matching templates for rule annotation, or by using a fine-tuned BERT-mini model for context-aware semantic classification, thereby transforming unstructured or weakly structured data into modelable semantic units.

[0053] Obtaining the feature association matrix for each semantically labeled category can be achieved by traversing the labeled data, counting the number of times any two labels co-occur in the same context, and normalizing the count. Furthermore, this operation can be implemented by dynamically updating the co-occurrence count based on a sliding time window, or by introducing mutual information or PMI indicators to replace the original frequencies to enhance sparsity robustness, thereby establishing a quantitative association view between semantic categories.

[0054] Constructing a community network model based on interaction strength weights using feature association matrices can be achieved by converting co-occurrence values ​​into weighted graph edges and calculating the final interaction strength weights by incorporating factors such as time decay and response differences. Furthermore, this operation can be achieved by using an exponential decay function to time-weight historical co-occurrences, or by adjusting the weights based on the variance of user group responses to highlight high signal-to-noise ratio associations, thereby constructing a weighted community network with business semantic meaning.

[0055] The segmented SMS data is divided into communities based on a community network model. The community segmentation results can be obtained by running a community detection algorithm (such as Louvain) on the community network model to divide semantic categories into multiple highly cohesive communities. Furthermore, this operation can be achieved by generating a hierarchical community structure using multi-resolution parameters or by combining node attribute embedding to enhance the semantic consistency of graph partitioning, thereby identifying semantically consistent and densely interacting subgroups.

[0056] A hybrid classification method based on industry scenario topology and user response decay paths utilizes community partitioning results to classify industry scenario features and user response decay features. This involves determining whether a data item within each community should be assigned to a scenario feature set or a decay feature set based on its semantic category's position in the industry scenario topology and the shape of its corresponding user decay path. Furthermore, this operation can be achieved by classifying semantic categories as industry scenario features if they are located on the topological backbone path and decay slowly, or by using a graph neural network to jointly encode the topological structure and decay path to output soft-classification probabilities, thus enabling accurate separation of the two types of features.

[0057] Taking a mixed scenario of quarterly bill notifications and marketing SMS messages in the financial industry as an example, the intelligent matching and effect prediction optimization method for SMS templates for industry scenarios in this embodiment can be as follows: At the end of the quarter, the bank simultaneously sends bill reminders and credit card limit increase marketing SMS messages. The system performs semantic annotation on the segmented data to obtain two categories of tags: "compliance notification" and "credit limit marketing"; constructing a feature association matrix reveals that the two frequently co-occur in some user groups but have low interaction intensity; after quantifying the interaction intensity weight, the community network model divides the two into independent communities; the hybrid classification method combines the financial scenario topology (compliance notifications are fixed nodes, marketing activities are branch nodes) and user decay path (the click-through rate of marketing content drops rapidly within 3 days while the notification remains stable), and finally classifies the bill data as target industry scenario data and the credit limit increase SMS data as target user response decay data, achieving accurate separation.

[0058] This embodiment forms semantically labeled categories by semantically annotating feature nodes of the segmented SMS data, constructs a feature association matrix reflecting the co-occurrence relationship between categories, builds a community network model using interaction intensity weights and performs community division, and finally reclassifies the community division results under dual constraints by a hybrid classification method that integrates industry scenario topology and user response decay path. This enables the data classification process to explicitly model scenario dynamics and user fatigue mechanisms, thereby improving the semantic purity and timeliness of feature extraction and providing high-quality structured data for subsequent time series analysis and prediction.

[0059] In one embodiment, scenario time-series evolution analysis is performed based on target industry scenario data to obtain scenario time-series evolution data, including: The fluctuation intensity of scene feature points is analyzed based on the target industry scene data to obtain the fluctuation intensity of scene feature points.

[0060] The fluctuation intensity of scene feature points can be a quantitative indicator measuring the magnitude of changes in the activity of semantic tags in a specific industry scene over time, and can be used to identify the intensity signal of sudden events or decaying trends. Furthermore, the acquisition method of the fluctuation intensity of scene feature points can be explained in context, i.e., it is achieved by calculating statistical fluctuation indicators (such as standard deviation, coefficient of variation, or local extrema). In this embodiment, the fluctuation intensity of scene feature points can include, but is not limited to, one or more of peak fluctuation intensity, variance fluctuation intensity, and relative growth rate fluctuation intensity. This operation can be performed through statistical analysis of the time series of each semantic tag in the target industry scene data. Further, this operation can use a sliding window to calculate local fluctuation intensity to capture instantaneous changes, or use wavelet transform to decompose high-frequency and low-frequency fluctuation components, thereby quantifying the severity of scene activity changes and identifying key event windows.

[0061] Based on target industry scenario data, curve fitting is used to analyze the evolution trajectory between scenario feature points, thereby obtaining the dynamic evolution trajectory between scenario feature points.

[0062] Curve fitting can be a modeling method that approximates discrete data points using mathematical functions to reveal their continuous change patterns. It can be used to characterize the smooth path and transition rate of the evolution of feature points in different scenarios over time. Furthermore, the operating principle of curve fitting can be explained in context, i.e., fitting multidimensional data points using parametric functions or interpolation methods. In this embodiment, curve fitting can employ one or more of polynomial fitting, spline interpolation fitting, and exponential decay fitting. Dynamic evolution trajectories can describe the continuous path and rate of change of semantic states of multiple industry scenarios in the time dimension, revealing the dynamic process of a scenario evolving from one business state to another. This operation can use the time series of multiple scenario feature points as coordinate points and fit a continuous function to describe their joint evolution path. Further, this operation can reveal the smooth transition patterns and evolution rates between semantic states of scenarios by performing principal component dimensionality reduction on multidimensional scenario features and then fitting a low-dimensional trajectory, or by using parametric spline curves to fit nonlinear evolution paths.

[0063] Based on the temporal transition probability network, the periodic extrapolation of scene time series is carried out by using fluctuation intensity and dynamic evolution trajectory to obtain periodic extrapolation data of scene time series.

[0064] The temporal transition probability network can be a directed graph model based on the Markov assumption, used to model the transition probabilities between scene states. It can be used to infer statistically significant periodic patterns in scenes. Furthermore, the acquisition method of the temporal transition probability network can be explained in context: using scene feature points as nodes, learning edge weights using historical transition frequencies, and adjusting the transition probabilities by weighting with fluctuation intensity. In this embodiment, the temporal transition probability network can be one or more of, including but not limited to, first-order Markov transition networks, higher-order temporal dependency networks, and hidden state transition networks. The periodic inference data of scene temporal sequence can be structured periodic pattern data representing the recurrence pattern of scenes, providing stable and reusable periodic context information for effect prediction. This operation can be to construct and infer the temporal transition probability network by using fluctuation intensity as state activation weights and dynamic evolution trajectory as transition direction constraints. Further, this operation can learn the hidden state transition matrix and infer the period using the EM algorithm, or use the autocorrelation function to initially screen the period length before constructing the transition network, thereby mining implicit, statistically significant periodic patterns from historical data.

[0065] Based on the fluctuation intensity and dynamic evolution trajectory, a deep time series prediction model is used to infer the scene evolution trend and obtain scene evolution trend inference data.

[0066] The deep time-series prediction model can be a machine learning model based on a neural network architecture that performs nonlinear modeling of multidimensional time-series features to predict future states. It can be used to make forward-looking inferences about the evolution trends of non-periodic and sudden scenarios. Furthermore, the operating principle of the deep time-series prediction model can be explained in context: it uses fluctuation intensity and dynamic evolution trajectory as input features, and learns long-term dependencies and mutation response mechanisms through end-to-end training. In this embodiment, the deep time-series prediction model can include, but is not limited to, one or more of the following: LSTM time-series prediction model, TemporalFusionTransformer, and N-BEATS model. The scenario evolution trend inference data can be a quantitative prediction result about the semantic state of the future scenario, which can enhance the system's adaptability to non-steady-state and non-periodic scenario changes. This operation can use the fluctuation intensity sequence and the derivative of the evolution trajectory as input features to train the deep time-series prediction model to make multi-step predictions of the future scenario state. Furthermore, this operation can enhance the robustness of trend inference by using an attention mechanism to dynamically weight the contribution of historical fluctuations to the future, or by introducing exogenous variables (such as holidays), thereby improving the ability to predict non-periodic and structural scenario changes.

[0067] Based on periodic extrapolation data and scenario evolution trend inference data, scenario temporal evolution analysis is performed to obtain scenario temporal evolution data.

[0068] The scene time-series evolution data can be a high-order scene time-series representation containing four-dimensional attributes—fluctuation, trajectory, periodicity, and trend—generated by fusing periodic projection data and trend inference data. This can serve as the core input for the effect-stage modeling. This operation can involve using periodic projection data as the steady-state component and trend inference data as the non-steady-state component, performing feature concatenation or weighted fusion. Furthermore, this operation can dynamically adjust the fusion weights of the periodic and trend components through a gating mechanism, or construct a dual-branch neural network to encode the periodicity and trend separately and then merge the outputs, thereby generating a high-dimensional scene time-series representation that combines periodic regularity and trend foresight. Taking the risk notification scenario during the quarterly financial report release period in the financial industry as an example, the intelligent matching and effect prediction optimization method for SMS templates for industry scenarios in this embodiment can be as follows: As the financial report season approaches, the system identifies feature points such as "financial report warning" and "compliance disclosure" from the target industry scenario data; fluctuation intensity analysis shows that the "compliance disclosure" tag starts to rise significantly 7 days before the financial report date; curve fitting reveals the S-shaped evolution trajectory of this feature point from low activity to high activity; the time series transition probability network deduces that the scenario has a quarterly periodicity, and the transition probability peaks 5 days before the financial report date; the deep time series prediction model further infers that the activity level in this quarter will increase by 20% compared with the same period in history due to new regulatory rules; finally, the system integrates the periodic and trend data to generate high-precision scenario time series evolution data, driving the risk control SMS template to be activated 3 days in advance, and adding the "new regulatory rule reminder" field, effectively improving the user open rate.

[0069] This embodiment analyzes the fluctuation intensity of scene feature points based on target industry scene data to obtain the fluctuation intensity of scene feature points; it uses curve fitting to analyze the evolution trajectory between scene feature points based on target industry scene data to obtain the dynamic evolution trajectory between scene feature points; it uses a time-series transition probability network to perform periodic extrapolation of scene time series based on fluctuation intensity and dynamic evolution trajectory to obtain periodic extrapolation data of scene time series; it uses a deep time series prediction model to infer scene evolution trend based on fluctuation intensity and dynamic evolution trajectory to obtain scene evolution trend inference data; and it performs scene time series evolution analysis based on periodic extrapolation data and scene evolution trend inference data to obtain scene time series evolution data. By quantifying the change amplitude of scene activity in multiple dimensions, revealing the semantic state transition law, mining periodic patterns, predicting non-periodic mutations, and integrating periodic and trend components, it can achieve the technical effect of overcoming the defects of static rule matching and extensive statistical prediction of traditional methods. This enables the SMS template matching mechanism to have the ability to perceive, understand and predict the dynamic evolution law of industry scenes, thereby supporting a full-cycle, adaptive intelligent optimization closed loop and significantly improving marketing response efficiency and resource utilization.

[0070] In one embodiment, SMS effect decay prediction is performed using effect stage modeling based on scenario time-series evolution data and target user response decay data to obtain effect decay prediction data, including: A response decay curve is generated based on the target user's response decay data, and the threshold intervals for the effect stages are divided based on the response decay curve.

[0071] The response decay curve can be a continuous function or discrete sequence reflecting the change in user response intensity to SMS over time or sending frequency, constructed based on target user response decay data. It can be used to visualize and quantify the user response decay trend, providing a basis for dividing the effect stages. Furthermore, the response decay curve can include, but is not limited to, one or more of the following: click-through rate decay curve, conversion rate decay curve, and unsubscription rate increase curve. The threshold range for the effect stage can be a segmentation boundary set on the response decay curve based on business semantics or statistical inflection points. This is used to divide the continuous decay process into discrete effect stages, enabling structured modeling of the SMS lifecycle (e.g., high response period, stable period, decay period). In this embodiment, the threshold range for the effect stage can be a high response threshold range, a stable maintenance threshold range, or a rapid decay threshold range, etc.

[0072] Generating response decay curves based on target user response decay data can be achieved by sorting the data by time or frequency and fitting a smooth curve or constructing an empirical distribution function. Furthermore, this operation can be implemented using nonparametric curve fitting with locally weighted regression (LOESS) or by modeling the decay trajectory with uncertainty using Gaussian process regression, thus transforming discrete user behavior into a continuous decay trend representation, facilitating stage segmentation and dynamic modeling. The threshold range for dividing the effect stages based on the response decay curve can be set according to changes in the curve slope, business rules, or clustering results, dividing the curve into several semantic stages. Further, this operation can be achieved by automatically identifying stage boundaries based on the extreme points of the first derivative or by introducing an absolute response rate threshold defined by business experts as the division criterion, thereby enabling discretized and interpretable modeling of the SMS marketing lifecycle.

[0073] Based on threshold intervals, migration analysis of the effect stage is performed to obtain effect stage migration data.

[0074] Specifically, effect stage migration data can be structured data describing the probability, rate, or triggering conditions of a user group transitioning from one effect stage to another. This data can be used to characterize the dynamic patterns of effect state evolution and support cross-stage extrapolation. Furthermore, effect stage migration data can be used in conjunction with scenario time-series evolution data for response decay extrapolation, enabling stage transition prediction to have external environment awareness capabilities.

[0075] Transition analysis of effect stages based on threshold intervals can be achieved by statistically analyzing the frequency and direction in which user groups cross threshold intervals within different time windows, constructing a state transition matrix or a Markov chain. Furthermore, this operation can be implemented by using hidden Markov models to infer potential stage transition paths, or by estimating the duration distribution of stays in each stage based on survival analysis. This allows for the quantification of the dynamic patterns of stage transitions, providing a transition mechanism for trend prediction.

[0076] Based on the effect stage migration data, the response decay is extrapolated using the scenario time-series evolution data to obtain response decay extrapolation data.

[0077] The response decay extrapolation data can be the result of extrapolating and simulating the future response decay path by combining the effect stage migration pattern and the scene time-series evolution data. It can be used to predict the future decay trajectory of the user response under a specific scene evolution path. Furthermore, the response decay extrapolation data can be one of the core inputs of the effect simulator, together with the noise robustness parameter, determining the final prediction result.

[0078] Using scenario-based time-series evolution data to extrapolate response decay based on effect stage transition data can be achieved by using scenario-based time-series evolution data (such as promotional periods or holidays) as external covariates to drive the effect stage transition model to simulate future paths. Furthermore, this operation can be implemented by constructing a recurrent neural network with exogenous inputs for sequence extrapolation, or by introducing scenario features as transition probability adjustment factors into the state transition model. This allows for the extrapolation of response decay trends across scenarios and time periods, improving predictive foresight.

[0079] Noise robustness is estimated based on scene time-series evolution data and target user response attenuation data to obtain noise robustness parameters.

[0080] Noise robustness estimation can be a process of modeling and suppressing anomalous disturbances in the temporal evolution data of the scene and the attenuation data of the target user response. This can be used to improve the stability and generalization ability of the performance prediction model in real-world complex data environments. Noise robustness parameters can be learnable or computable parameters that characterize the level of data noise and its impact on the prediction results. These can be used to adjust the uncertainty weights during performance simulation, enhancing the anti-interference ability of the prediction results. Furthermore, noise robustness parameters can be estimated using robust statistical or variational inference methods by jointly modeling scene fluctuations and user behavior outliers.

[0081] Noise robustness estimation based on scene time-series evolution data and target user response decay data can involve jointly analyzing anomalous patterns (such as sudden clicks or system delays) in both types of data to estimate noise distribution or uncertainty weights. Furthermore, this operation can be achieved by training the model using the Huber loss function and extracting residual distribution parameters, or by using a variational autoencoder (VAE) to separate the latent representations of signal and noise. This enhances the model's tolerance to real-world data perturbations and avoids overfitting and accidental behavior.

[0082] Based on response decay extrapolation data and noise robustness parameters, the effect simulator is used to predict SMS effect decay and obtain effect decay prediction data.

[0083] The effect simulator can be a computational module that integrates response decay projection data with noise robustness parameters to simulate the evolution of SMS effects in future dynamic scenarios. It can be used to generate high-confidence, interference-resistant effect decay prediction data, supporting refined risk prediction. Furthermore, the effect simulator can construct dynamic response evolution models based on stochastic processes, Monte Carlo simulations, or differential equation systems, and can include, but is not limited to, deterministic effect simulators, stochastic perturbation effect simulators, and multi-scenario branch simulators.

[0084] Predicting SMS effect decay using an effect simulator based on response decay projection data and noise robustness parameters can be achieved by injecting response decay projection paths into the effect simulator, adjusting the intensity of random perturbations in the simulation according to noise robustness parameters, and outputting a statistical aggregation of multi-round simulation results. Furthermore, this operation can be achieved by generating multiple possible decay paths through Monte Carlo simulation and then taking quantiles, or by modeling the response decay rate using a differential equation system and adding random perturbation terms for solution, thereby generating high-confidence prediction results that combine trend accuracy and noise resistance stability.

[0085] Taking the optimization of monthly bill reminder SMS messages in the financial industry as an example, the intelligent matching and effect prediction optimization method for SMS templates for industry scenarios in this embodiment can be as follows: At the beginning of each month, the system generates a response decay curve based on the historical user ignore rate and unsubscribe behavior of bill SMS messages, identifying two stages: "high response on the first day" and "rapid decay after three days"; through effect stage migration analysis, it is found that if it is a holiday, the speed at which users enter the decay period is significantly accelerated; combined with the scene time-series evolution data of the monthly calendar, the decay acceleration trend of this bill cycle is deduced; at the same time, noise robustness estimation identifies that some users have abnormally low SMS clicks due to APP push interference, and the noise parameters are adjusted accordingly; the effect simulator integrates the above information and predicts that if it is sent at 10 am on a weekday, the high response period can be extended by about 8 hours; this prediction result is used to generate an optimization plan of "avoiding weekends + sending in the morning", which effectively improves the open rate.

[0086] This embodiment generates a response decay curve based on target user response decay data and divides the effect stages into threshold intervals based on the response decay curve, achieving continuous modeling and discretized stage division of the user response decay trend. It obtains effect stage migration data that quantifies the stage transition rules by performing effect stage migration analysis based on the threshold intervals. It obtains future decay trajectory predictions across scenarios by combining scenario time-series evolution data to drive response decay extrapolation. It obtains noise robustness parameters that enhance anti-interference capabilities by jointly estimating noise robustness using scenario and user data. Finally, it generates noise-resistant and stable prediction results by fusing response decay extrapolation data and noise robustness parameters into an effect simulator. This process, through the synergistic effect of structured modeling, dynamic migration analysis, scenario-driven extrapolation, and noise suppression, overcomes the limitations of traditional prediction methods, achieving comprehensive and accurate prediction of SMS effect decay from trend representation to dynamic extrapolation, and from single-dimensional prediction to multi-scenario adaptation. This significantly improves the timeliness, anti-interference capability, and strategy adaptation accuracy of the prediction results.

[0087] In one embodiment, an industry knowledge graph is constructed. Based on scenario time-series evolution data and effect decay prediction data, the industry knowledge graph is used to predict the potential effect risk of SMS templates, obtaining potential effect risk data, including: The system acquires historical performance data and industry scenario topology network for SMS templates. The historical performance data includes historical scenario data and historical response data. Based on the historical performance data, the system performs scenario semantic level division and response quality level division to obtain scenario semantic level division data and response quality level division data.

[0088] The historical performance data of SMS templates can be a collection of sending records and user feedback results of SMS templates in different scenarios within a historical period. This data can be used as the basic training and structuring basis for building a risk identification model. Furthermore, historical scenario data can be meta-information related to the business context in the historical performance data, such as activity type, time window, and industry tags. This data can be used to support the semantic hierarchy of scenarios and reflect the granularity of event abstraction. Historical response data can be user behavior feedback indicators for SMS messages in the historical performance data, such as click-through rate, unsubscription rate, and conversion rate. This data can be used to classify response quality levels and characterize the intensity of user feedback. The industry scenario topology network can be graph structure data describing the semantic associations, temporal dependencies, or co-occurrence relationships between different industry scenarios. This data can be used to provide the inter-scenario association context for the performance risk matrix, enhancing the structure of risk modeling. In this embodiment, the industry scenario topology network can be constructed by using log co-occurrence analysis, expert rules, or entity linking technology to build scenario nodes and their edge relationships. For example, the industry scenario topology network can include, but is not limited to, one or more of the following: e-commerce scenario association network, financial notification dependency network, and logistics service trigger network.

[0089] Scene semantic hierarchy classification data can be the result of multi-level classification of historical scene data according to the degree of semantic abstraction, and can be used to establish a hierarchical representation from specific events to general categories. Furthermore, scene semantic hierarchy classification data can include, but is not limited to, event-level scenes, activity-level scenes, and domain-level scenes. Response quality hierarchy classification data can be the result of discretizing and classifying historical response data according to the intensity of user feedback, and can be used to transform continuous response indicators into interpretable quality levels, facilitating risk modeling. Furthermore, response quality hierarchy classification data can include, but is not limited to, high response level, medium response level, and low response level.

[0090] Obtaining historical performance data and industry scenario topology networks for SMS templates can be achieved by extracting template-level performance logs from historical databases and loading pre-built scenario association graphs. Furthermore, this operation can be implemented by batch synchronizing historical performance data to a feature repository via an ETL pipeline, or by real-time calling a graph database API to obtain the latest industry scenario topology network, thus providing a structured input source for subsequent hierarchical partitioning and risk modeling. Scene semantic hierarchical partitioning and response quality hierarchical partitioning based on historical performance data can be achieved by applying ontology classification rules to historical scenario data for upward abstraction, and by discretely grading historical response data according to quantiles or clustering results. Furthermore, this operation can be achieved by using an industry ontology library (such as e-commerce OWL) for automatic semantic classification, or by using K-means to perform tri-cluster partitioning of the response rate distribution to define quality levels, thereby transforming the raw data into a structured representation with semantic hierarchy and quality levels.

[0091] A risk matrix for the effect of constructing data based on scene semantic hierarchy classification data, response quality hierarchy classification data, and industry scene topology network is built.

[0092] The effect-risk matrix can be a two-dimensional structured table with scene semantic level as rows and response quality level as columns, quantifying the probability of SMS failure under each combination. It can provide a preliminary risk quantification view as the input basis for constructing risk chain relationships. Furthermore, the effect-risk matrix can be based on the failure frequency of each cell calculated from historical effect data, and then smoothed or weighted by integrating an industry scene topology network. In this embodiment, the effect-risk matrix can be combined with neighboring scenes in the industry scene topology network for Bayesian smoothing, or graph attention weights can be introduced to weight and fuse the risk values ​​of adjacent scenes.

[0093] The effect risk matrix, constructed based on scenario semantic hierarchy data, response quality hierarchy data, and industry scenario topology network, can be created by statistically analyzing the historical failure rate of each cross-unit, with scenario hierarchy as rows and response hierarchy as columns. Furthermore, this operation can be achieved by using Laplace smoothing to process sparse cells, thereby generating a preliminary structured risk quantification view that reflects the failure tendency of different scenario-response combinations.

[0094] Based on historical performance data, an effect-risk coupling degree analysis is performed to obtain effect-risk coupling degree data.

[0095] Among them, the effect-risk coupling degree data can be a quantitative indicator reflecting the intensity of the interaction between multiple risk factors (such as content duplication, non-prime time periods, and high transmission frequency). It can be used to reveal complex risk patterns beyond single factors and improve the nonlinear perception capability of risk identification. Furthermore, the effect-risk coupling degree data can be used to analyze the combined impact of multiple factors on response attenuation through mutual information, Shapley values, or causal inference methods.

[0096] Analysis of the coupling degree between effects and risks based on historical effect data can be conducted by identifying the joint occurrence patterns of multiple risk factors and calculating the strength of their synergistic impact on response attenuation. Furthermore, this operation can be achieved by quantifying the synergistic effects between factors using SHAP interaction values, or by constructing a causal forest model to estimate response changes under multi-factor interventions, thereby revealing complex risk effects and going beyond the assumption of single-factor linear superposition.

[0097] Risk chain relationships are constructed based on the effect-risk matrix and effect-risk coupling degree data to obtain a risk chain structure diagram, and an industry knowledge graph is constructed based on the risk chain structure diagram.

[0098] Among them, the risk chain structure graph can be a directed graph representing the transmission paths and causal dependencies between risk factors. It can be used to explicitly model chain failure mechanisms such as "template duplication, user fatigue, and response decline." Furthermore, the risk chain structure graph can include, but is not limited to, content fatigue transmission chains, timing mismatch transmission chains, and scenario misalignment transmission chains. Industry knowledge graphs can be structured knowledge bases that can embed dynamic risk logic, and can be used to support intelligent reasoning.

[0099] Constructing risk chain relationships based on effect-risk matrices and effect-risk coupling data can be achieved by using high-risk matrix units as starting nodes, deriving transmission paths using coupling data, and constructing directed edges to represent causal or temporal dependencies. Furthermore, this operation can determine the temporal transmission direction based on Granger causality tests, or utilize a rule engine to encode expert experience into chain-based construction constraints, thereby explicitly modeling risk transmission mechanisms and forming interpretable failure chains. Constructing an industry knowledge graph based on a risk chain structure graph can be achieved by expanding the nodes and edges in the risk chain structure graph into a triplet knowledge graph containing attributes, types, and relationship types. Furthermore, this operation can be achieved by converting the chain graph to RDF format and injecting it into a graph database, or by adding "risk_cause" and "risk_effect" relationship types to the existing industry knowledge graph, thus enabling the knowledge graph to not only contain static entities but also embed dynamic risk logic.

[0100] A potential effect risk prediction model is constructed based on industry knowledge graphs and graph neural network prediction models. This model is then used to predict the potential effect risk of SMS templates using scenario time-series evolution data and effect decay prediction data, thus obtaining potential effect risk data. The graph neural network prediction model can be a deep neural network architecture capable of learning node representations and reasoning about relationships in graph-structured data. It can be used to empower industry knowledge graphs to achieve high-order semantic reasoning and capture non-local risk associations. Furthermore, the graph neural network prediction model can include, but is not limited to, one or more of the following: GraphSAGE, GAT, and R-GCN models. The potential effect risk prediction model can be an end-to-end risk prediction system jointly constructed from industry knowledge graphs and graph neural network prediction models. It can be used to achieve dynamic and accurate prediction of the potential failure risk of SMS templates in new scenarios.

[0101] Constructing a potential effect risk prediction model based on an industry knowledge graph combined with a graph neural network prediction model can be achieved by using the industry knowledge graph as the input graph structure, loading a graph neural network model for end-to-end training, and embedding risk semantics into nodes. Furthermore, this operation can enhance inference accuracy by using TransR to map relation types to specific subspaces, or by using graph contrastive learning to pre-train node representations to improve robustness in small-sample scenarios, thereby achieving the ability to generalize and predict risks for unseen template-scenario combinations in the graph. Predicting the potential effect risk of SMS templates using scenario temporal evolution data and effect decay prediction data can be achieved by mapping the current scenario temporal evolution data and effect decay prediction data to query nodes or attributes in the graph, and outputting a risk score through graph neural network inference. Furthermore, this operation can be achieved by dynamically updating the temporal evolution features as node attributes before performing inference, or by matching similar historical scenarios through graph subgraph retrieval and then performing risk migration prediction, thereby achieving dynamic and forward-looking prediction of template risks in new scenarios, overcoming the limitations of static statistics.

[0102] Taking the risk prediction of collection SMS messages in the financial industry before and after the repayment date as an example, the intelligent matching and effect prediction optimization method of SMS templates for industry scenarios in this embodiment can be as follows: The system obtains the historical effect data of collection SMS messages of a certain bank over the past 6 months and the financial scenario topology network (such as "overdue reminder" often co-occurring with "repayment guide"). By performing semantic hierarchical division of historical scenario data, it is found that "M1 overdue collection" belongs to the "short-term collection" subcategory; the response data is divided into three layers: high (clicks > 5%), medium (1%-5%), and low (< 1%). The constructed effect risk matrix shows that the failure rate of the "short-term collection + low response" combination is as high as 78%. Coupling analysis found that the risk coupling degree of the combination of "sending on weekends + no personalized content" is significantly higher than that of individual factors. Based on this, a risk chain structure diagram is constructed: "sending on non-working days, low open rate, lack of repayment action, and increased bad debt risk". This chain is integrated into the financial industry knowledge graph. When a new collection plan is launched, the potential effectiveness risk prediction model combines the current scenario's temporal evolution (increased fluctuations as the repayment date approaches) with the effectiveness decay prediction (the user has received 3 similar text messages). It predicts that the current template has a high risk and suggests switching to the "weekday morning + personalized debt details" strategy to avoid ineffective outreach.

[0103] This embodiment acquires historical performance data of SMS templates and industry scenario topology networks. Based on the historical performance data, it performs scenario semantic hierarchy division and response quality hierarchy division, constructs an performance risk matrix, analyzes performance risk coupling degree based on historical performance data, constructs risk chain relationships based on the performance risk matrix and performance risk coupling degree data, and forms an industry knowledge graph. It then combines a graph neural network prediction model to construct a potential performance risk prediction model, and uses scenario time-series evolution data and performance decay prediction data for prediction. By transforming raw data into a structured hierarchical representation, quantifying compound risk effects, explicitly modeling risk transmission paths, and integrating knowledge graphs and deep learning reasoning, it can systematically solve the problems of traditional methods relying on manual rules and lacking semantic depth. It achieves dynamic and accurate prediction of potential failure risks of SMS templates in new scenarios, significantly improves prediction foresight and strategy adaptation accuracy, supports the generation of targeted optimization plans, effectively alleviates user response fatigue, improves marketing conversion efficiency, and promotes the evolution of marketing systems into intelligent agents with cognitive capabilities.

[0104] In one embodiment, semantic association matching is performed on potential effect risk data and various SMS optimization strategies to obtain semantic association matching results, including: Obtain the semantic feature vector of potential effect risk data and the strategy feature vector of each SMS optimization strategy.

[0105] The semantic feature vector can be a dense vector representation obtained by semantically encoding the potential effect risk data, used to characterize the semantic connotation of the risk in the embedding space. It is understood that the semantic feature vector can be generated by encoding the risk description text using a pre-trained language model (such as BERT) or a graph neural network. Furthermore, in this embodiment, the semantic feature vector, as the basic representation for semantic association matching, can be used to support the calculation of similarity and difference with the strategy feature vector. The strategy feature vector can be an embedded representation of the semantic description of the SMS optimization strategy after vectorization, and can be used for multi-dimensional matching evaluation with the semantic feature vector, supporting the intelligent association of strategy-risk pairs. It is understood that the strategy feature vector can be obtained using the same encoder or alignment training mechanism as the semantic feature vector, ensuring that it is in a unified semantic space. In this embodiment, obtaining the semantic feature vector of the potential effect risk data and the strategy feature vector of each SMS optimization strategy can be achieved by calling a unified semantic encoder to vectorize the risk description text and the strategy description text respectively. Furthermore, this operation can be achieved by using a dual-tower BERT model with shared parameters to encode risk and policy texts separately, or by generating structure-aware feature vectors based on a graph embedding model pre-trained on an industry knowledge graph. This allows for the establishment of comparable representations of risk and policy in the same semantic space, providing a foundation for subsequent matching.

[0106] The semantic graph neural network is used to measure the similarity of semantic feature vectors and feature vectors of each strategy, and the corresponding similarity measurement results are obtained.

[0107] Semantic graph neural networks (SBRs) can be neural network models that integrate semantic information with graph structure relationships, used to propagate and aggregate higher-order semantic dependencies between nodes. Understandably, the operating principle of a SBR can be to construct a risk-policy bipartite graph using semantic feature vectors and policy feature vectors as graph nodes, updating node representations and calculating similarity through a multi-layer message passing mechanism. Measuring the similarity of semantic feature vectors and policy feature vectors based on SBRs can involve constructing a risk-policy interaction graph, using the SBR for multi-hop information propagation, and calculating node pair similarity. Furthermore, this operation can be achieved by using a GAT layer to aggregate neighboring policy node information and then calculating attention-weighted similarity, or by introducing a contrastive learning loss at the end of the graph neural network to shorten the distance between positive sample pairs. This allows for capturing higher-order semantic associations between risk and policy, surpassing shallow similarity calculations such as dot product or cosine similarity. The similarity measurement result can be a numerical indicator reflecting the semantic similarity between semantic feature vectors and policy feature vectors, used to measure the degree to which the policy semantically covers risk requirements, serving as a one-dimensional basis for matching decisions.

[0108] Obtain the difference feature set between the semantic feature vector and the feature vector of each strategy.

[0109] The difference feature set can be a set of feature dimensions extracted from the comparison of semantic feature vectors and policy feature vectors, representing semantic inconsistencies or capability gaps between the two. Understandably, the difference feature set can be obtained through vector differencing, attention weight analysis, or adversarial discriminators to identify key divergence dimensions. Furthermore, obtaining the difference feature set between semantic feature vectors and each policy feature vector can be achieved by analyzing the directional deviation or dimensional activation differences between the two vectors in the embedding space to extract significant divergence features. This operation can be further implemented by calculating the dimension-wise residuals and selecting dimensions with an L1 norm greater than a threshold as difference features, or by training an interpretable difference detector (such as a sparse linear probe) to locate key divergence points, thereby identifying risky semantic dimensions not covered by the policy and providing input for difference assessment.

[0110] Based on the difference feature set, heterogeneous graph matching is used to evaluate the difference between semantic feature vectors and feature vectors of each strategy, and the corresponding difference evaluation results are obtained.

[0111] Heterogeneous graph matching can be a subgraph or node pair matching algorithm designed for graph structures containing different types of nodes and edges. Understandably, the operating principle of heterogeneous graph matching can be to treat semantic feature vectors and policy feature vectors as different node types in a heterogeneous graph, construct attributed edges based on a set of difference features, and calculate the difference degree through a graph matching network. Evaluating the difference degree using heterogeneous graph matching based on the set of difference features can involve constructing a heterogeneous graph using difference features as edge attributes or node modifiers, and running the graph matching network to output a difference score. Furthermore, this operation can be achieved by using a GMN architecture to structurally align the risk-policy subgraph and output difference embeddings, or by constructing a dual-channel heterogeneous graph: one channel representing commonalities and the other focusing on differences, jointly outputting the difference degree. This quantifies the degree of policy deficiency in key risk dimensions, preventing the recommendation of highly similar but practically inapplicable policies. The difference degree evaluation result can be an indicator that quantifies the degree of inconsistency between semantic feature vectors and policy feature vectors in key dimensions, serving as a supplementary basis for matching decisions.

[0112] Based on the similarity quantification results and the difference assessment results, semantic association matching was performed on the potential effect risk data and each SMS optimization strategy to obtain semantic association matching results.

[0113] Semantic association matching based on similarity metric results and difference assessment results can be achieved by fusing similarity scores and difference scores (e.g., weighted summation, gated fusion, or ranking fusion) to generate a comprehensive matching score. Furthermore, this operation can dynamically balance the contributions of similarity and difference through learnable weights, or use both as input features for a multi-objective ranking model, optimizing matching quality end-to-end. This allows for strategy recommendation that balances semantic fit and capability completeness, improving matching accuracy and interpretability. Taking the increase in user unsubscription rates due to sensitive content in financial risk control notification SMS messages as an example, the intelligent matching and effect prediction optimization method for SMS templates in this embodiment, tailored to industry scenarios, can be as follows: The system identifies a certain type of risk control SMS message as having a potential risk of "decreased user trust," and its semantic feature vector emphasizes "insufficient authority" and "vague explanation." In the strategy library, the strategy feature vector of the "adding endorsement language from regulatory agencies" strategy has a high similarity, but the difference feature set shows that this strategy does not cover the "lack of operational guidance" dimension. Heterogeneous graph matching gives a moderate difference score accordingly. Although the final comprehensive matching score is not the highest, it is still included in the plan because the difference is controllable. Another strategy, "shortening the text length," has a slightly higher similarity but an extremely high difference (completely does not involve trust building) and is effectively filtered out. The system thus generates a composite optimization strategy that includes authoritative language and operational guidance, significantly reducing the unsubscription rate.

[0114] This embodiment establishes a comparable representation of risk and strategy in the same semantic space by acquiring the semantic feature vectors of potential effect risk data and the strategy feature vectors of each SMS optimization strategy. It captures the high-order semantic relationship between risk and strategy by performing similarity metric on the semantic feature vectors and strategy feature vectors based on a semantic graph neural network. It identifies risk semantic dimensions not covered by the strategies by acquiring a difference feature set. It quantifies the degree of strategy deficiency in key dimensions by using heterogeneous graph matching based on the difference feature set. Finally, it achieves comprehensive matching by fusing the similarity metric results and the difference assessment results, ultimately realizing strategy recommendation that balances semantic fit and capability completeness. This significantly improves the accuracy, robustness, and interpretability of strategy recommendation, enabling optimization strategies to truly fit the fine-grained risk characteristics in dynamic scenarios, supporting the generation quality of effect risk management plans, and thus enhancing the technical effect of targeted and effective dynamic optimization of SMS templates.

[0115] Furthermore, to achieve the above objectives, the present invention also provides an intelligent matching and effect prediction optimization device for SMS templates in industry scenarios. The device includes: a memory, a processor, and an intelligent matching and effect prediction optimization program for SMS templates in industry scenarios stored on the memory and executable on the processor. The intelligent matching and effect prediction optimization program for SMS templates in industry scenarios is configured to implement the steps of the intelligent matching and effect prediction optimization method for SMS templates in industry scenarios as described above.

[0116] In addition, to achieve the above objectives, the present invention also provides a medium storing an industry-specific intelligent matching and effect prediction optimization program for SMS templates, wherein when the industry-specific intelligent matching and effect prediction optimization program is executed by a processor, it implements the steps of the industry-specific intelligent matching and effect prediction optimization method for SMS templates as described above.

[0117] Other embodiments or specific implementations of the intelligent matching and effect prediction optimization device for SMS templates for industry scenarios described in this invention can be referred to the above-mentioned method embodiments, and will not be repeated here.

[0118] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios, characterized in that, The method is applied to an SMS service platform, which is equipped with a machine learning inference engine; the method includes: The machine learning inference engine of the SMS service platform uses a data collection strategy to obtain SMS sending records and user interaction behavior data, and preprocesses the data to obtain preprocessed SMS data. The preprocessed SMS data is divided into blocks to obtain the block-processed SMS data; based on the community detection algorithm, the block-processed SMS data is classified according to industry scenario characteristics and user response decay characteristics to obtain target industry scenario data and target user response decay data. Based on the target industry scenario data, scenario time-series evolution analysis is performed to obtain scenario time-series evolution data; based on the scenario time-series evolution data and target user response decay data, effect stage modeling is used to predict SMS effect decay and obtain effect decay prediction data. Construct an industry knowledge graph, and use the industry knowledge graph to predict the potential effect risk of SMS templates based on the scenario time-series evolution data and effect decay prediction data to obtain potential effect risk data. Several SMS optimization strategy libraries are obtained from the database. Semantic association matching is performed on the potential effect risk data and each SMS optimization strategy to obtain semantic association matching results. An effect risk management plan is generated based on the semantic association matching results. The SMS template is dynamically optimized based on the effect risk management plan.

2. The method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios as described in claim 1, characterized in that, The process of acquiring SMS sending records and user interaction behavior data using a data collection strategy, and preprocessing the data to obtain preprocessed SMS data, includes: The data collection cycle and data collection granularity are determined based on preset industry labels, and the data collection threshold is constructed based on the dynamic coupling relationship between the data collection cycle and data collection granularity. A data collection strategy is determined based on the data collection threshold, and SMS sending records and user interaction behavior data are obtained based on the data collection strategy. The SMS data is subjected to noise filtering to obtain noise-filtered SMS data. Feature alignment is performed on the noise-filtered SMS data to obtain preprocessed SMS data.

3. The method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios as described in claim 1, characterized in that, The step of dividing the preprocessed SMS data into blocks to obtain block-processed SMS data includes: A dynamic time window is set, and the preprocessed SMS data is divided into blocks based on the continuity of scene boundaries to obtain the block-processed SMS data.

4. The method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios as described in claim 1, characterized in that, The method involves classifying the segmented SMS data based on community detection algorithms according to industry scenario characteristics and user response decay characteristics to obtain target industry scenario data and target user response decay data, including: The feature nodes of the segmented SMS data are semantically annotated to obtain several semantic annotation categories; Obtain the feature association matrix for each semantic labeling category, and construct a community network model based on the feature association matrix using interaction strength weight quantization. Based on the community network model, the segmented SMS data is divided into communities to obtain the community division results. A hybrid classification method based on industry scenario topology and user response decay path is used to classify data on industry scenario characteristics and user response decay characteristics using the community segmentation results, thereby obtaining target industry scenario data and target user response decay data.

5. The method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios as described in claim 1, characterized in that, The step of performing scenario time-series evolution analysis based on the target industry scenario data to obtain scenario time-series evolution data includes: Based on the target industry scenario data, the fluctuation intensity of scenario feature points is analyzed to obtain the fluctuation intensity of scenario feature points; Based on the target industry scenario data, curve fitting is used to analyze the evolution trajectory between scenario feature points to obtain the dynamic evolution trajectory between scenario feature points. Based on the temporal transition probability network, the fluctuation intensity and dynamic evolution trajectory are used to periodically extrapolate the temporal sequence of the scene, thereby obtaining periodic extrapolation data of the temporal sequence of the scene. Based on the fluctuation intensity and dynamic evolution trajectory, a deep time series prediction model is used to infer the scene evolution trend and obtain scene evolution trend inference data. Based on the periodic extrapolation data and the scenario evolution trend inference data, scenario temporal evolution analysis is performed to obtain scenario temporal evolution data.

6. The method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios as described in claim 1, characterized in that, The method of predicting SMS effect decay by using effect stage modeling based on the scenario time-series evolution data and target user response decay data to obtain effect decay prediction data includes: A response decay curve is generated based on the target user response decay data, and threshold intervals for the effect stages are divided based on the response decay curve. Based on the threshold range, perform migration analysis of the effect stage to obtain effect stage migration data; Based on the effect stage migration data, the response decay is extrapolated using the scenario time-series evolution data to obtain response decay extrapolation data. Noise robustness is estimated based on the time-series evolution data of the scenario and the target user response attenuation data to obtain noise robustness parameters; Based on the response attenuation extrapolation data and noise robustness parameters, the SMS effect attenuation is predicted using an effect simulator to obtain effect attenuation prediction data.

7. The method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios as described in claim 1, characterized in that, The construction of the industry knowledge graph involves using the scenario time-series evolution data and effect decay prediction data to predict the potential effect risk of SMS templates, thereby obtaining potential effect risk data, including: The historical effect data and industry scenario topology network of SMS templates are obtained. The historical effect data includes historical scenario data and historical response data. Based on the historical effect data, scenario semantic level division and response quality level division are performed to obtain scenario semantic level division data and response quality level division data. Based on the aforementioned scenario semantic hierarchy segmentation data, response quality hierarchy segmentation data, and industry scenario topology network, an effect risk matrix is ​​constructed. Based on the historical performance data, an effect-risk coupling degree analysis is performed to obtain effect-risk coupling degree data; Based on the effect-risk matrix and effect-risk coupling data, a risk chain relationship is constructed to obtain a risk chain structure diagram, and an industry knowledge graph is constructed based on the risk chain structure diagram. A potential effect risk prediction model is constructed based on the industry knowledge graph and graph neural network prediction model. Based on the potential effect risk prediction model, the potential effect risk of SMS templates is predicted using the scenario time-series evolution data and effect decay prediction data to obtain potential effect risk data.

8. The method for intelligent matching and effect prediction optimization of SMS templates for industry scenarios as described in claim 1, characterized in that, The step of performing semantic association matching on the potential effect risk data and each SMS optimization strategy to obtain semantic association matching results includes: Obtain the semantic feature vector of the potential effect risk data and the strategy feature vector of each SMS optimization strategy; The semantic feature vector and each strategy feature vector are similarity quantified based on the semantic graph neural network to obtain the corresponding similarity quantification results. Obtain the difference feature set between the semantic feature vector and the feature vectors of each strategy; Based on the aforementioned difference feature set, heterogeneous graph matching is used to evaluate the difference between semantic feature vectors and feature vectors of each strategy, and the corresponding difference evaluation results are obtained. Based on the similarity quantification results and the difference assessment results, semantic association matching is performed on the potential effect risk data and each SMS optimization strategy to obtain semantic association matching results.

9. A device for intelligent matching and effect prediction optimization of SMS templates for industry scenarios, characterized in that, The device includes: a memory, a processor, and an industry-specific intelligent matching and effect prediction optimization program for SMS templates stored in the memory and executable on the processor. The industry-specific intelligent matching and effect prediction optimization program for SMS templates is configured to implement the steps of the industry-specific intelligent matching and effect prediction optimization method for SMS templates as described in any one of claims 1 to 8.

10. A medium, characterized in that, The medium stores an intelligent matching and effect prediction optimization program for SMS templates oriented to industry scenarios. When the processor executes the intelligent matching and effect prediction optimization program for SMS templates oriented to industry scenarios, it implements the steps of the intelligent matching and effect prediction optimization method for SMS templates oriented to industry scenarios as described in any one of claims 1 to 8.

Images