Intelligent robot-assisted short message push management system and method

Through the intelligent robot-assisted SMS push management system, a personalized push solution is generated using multi-scale feature extraction and feature fusion analysis, which solves the problem that traditional SMS push cannot meet users' individual needs and improves push effect and user satisfaction.

CN118827775BActive Publication Date: 2024-11-29XIAN SANLUJIUJIU NETWORK TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411299746.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-18
Publication Date
2024-11-29
Estimated Expiration
2044-09-18

AI Technical Summary

Technical Problem

Traditional SMS push usually adopts a common push strategy, which fails to fully combine users' individual differences and behavioral patterns, resulting in the inability to meet users' individual needs and affect the push effect.

Method used

Provide intelligent robot-assisted SMS push management system and method, and generate user personalized push solutions through push feature collection, historical behavior data acquisition, multi-scale feature extraction, framework sequence construction, in-sequence adjacency matrix fusion analysis and feature fusion analysis, and generate user personalized push solutions and optimize Push time nodes to realize personalized push management.

Benefits of technology

It improves the response rate and user satisfaction of push, ensures that push content can be sent at the time that users are most likely to accept, meets user needs and preferences, and improves push effect and user participation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118827775B_ABST
    Figure CN118827775B_ABST
Patent Text Reader

Abstract

The present invention provides an intelligent robot-assisted SMS push management system and method, which relates to the technical field of SMS push management, including: collecting a push feature set of a target push service; obtaining a user historical behavior data sequence set; using K adaptive convolutional network layers to perform multi-scale feature extraction on historical push feedback and service usage data to construct a three-dimensional framework sequence; performing adjacency matrix fusion analysis to obtain a mapping feature sequence of feedback and service usage; performing feature fusion analysis through a fully connected layer, reordering the user feature sequence, determining a personalized push time node, and generating a personalized push plan. The invention solves the technical problem that traditional SMS push usually adopts a general push strategy, fails to fully combine the individual differences and behavior patterns of users, and leads to poor push effects. Through in-depth analysis of user behavior patterns, personalized push management is realized, thereby achieving the technical effect of improving push effects.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of SMS push management, and in particular to an intelligent robot-assisted SMS push management system and method. Background Art

[0002] Accurate SMS push can significantly improve marketing effectiveness and customer satisfaction. However, traditional SMS push usually formulates push strategies based on simple rules, such as users' basic information or historical purchase records. This approach ignores the dynamic changes in user behavior. Even for the same customer, the degree of acceptance of the same service may be different in different time periods. This method cannot adapt to changes in user needs, nor can it use users' specific behavior data in different time periods to optimize push content. This leads to the possibility that the push content may not match the user's interests and needs, thus affecting the push effect. Summary of the invention

[0003] This application provides an intelligent robot-assisted SMS push management system and method, aiming to solve the technical problem that traditional SMS push usually adopts a universal push strategy, fails to fully consider the individual differences and behavior patterns of users, and cannot meet the individual needs of users, resulting in poor push effect.

[0004] The first aspect disclosed in the present application provides an intelligent robot-assisted SMS push management system, the system comprising: a push feature acquisition module, the push feature acquisition module is used to perform push feature acquisition and obtain a push feature set of a target push service; a historical behavior data acquisition module, the historical behavior data acquisition module is used to interact with an intelligent robot to obtain a pre-collected historical behavior data sequence set of a target push user set of the target push service, wherein the historical behavior data sequence set includes a historical push feedback sub-data sequence set and a historical service usage sub-data sequence set of the target push user set within a preset historical acquisition window; a multi-scale feature acquisition module; a historical behavior data ... A feature extraction module, the multi-scale feature extraction module is used to use K adaptive convolutional network layers to perform multi-scale feature extraction on the historical push feedback sub-data sequence set and the historical business use sub-data sequence set according to K receptive fields, to obtain K historical push feedback sub-data feature cluster sequence sets and K historical business use sub-data feature sequence cluster sets, wherein the sizes of the K receptive fields are inconsistent; a frame sequence acquisition module, the frame sequence acquisition module is used to construct an initial three-dimensional frame sequence, and input the K historical push feedback sub-data feature cluster sequence sets and the K historical business use sub-data feature cluster sequence sets into the initial three a frame sequence, generating K historical push feedback frame sequence sets and K historical business usage frame sequence sets; an adjacency matrix fusion analysis module, the adjacency matrix fusion analysis module is used to perform intra-sequence adjacency matrix fusion analysis on the K historical push feedback frame sequence sets and the K historical business usage frame sequence sets, and obtain K historical push feedback mapping feature sequence sets and K historical business usage mapping feature sequence sets; a feature fusion analysis module, the feature fusion analysis module is used to transmit the K historical push feedback mapping feature sequence sets and the K historical business usage mapping feature sequence sets to the fully connected layer for feature fusion analysis, Obtain a user fusion feature sequence set; a push time node determination module, the push time node determination module is used to reorder the user fusion feature sequence set according to acceptance from large to small, obtain a user push fusion feature sequence set, and determine a user push time node sequence set based on the user push fusion feature sequence set; an SMS push management module, the SMS push management module is used to identify the user personalized push plan of the target push service based on the user push time node sequence set, obtain the user personalized push plan set, and perform SMS push management of the target push service based on the user personalized push plan set.

[0005] The second aspect disclosed in the present application provides an intelligent robot-assisted SMS push management method, which is implemented by the above-mentioned intelligent robot-assisted SMS push management system, and the method includes: executing push feature collection to obtain a push feature set of a target push service; the interactive intelligent robot obtains a pre-collected historical behavior data sequence set of a target push user set of the target push service, wherein the historical behavior data sequence set includes a historical push feedback sub-data sequence set and a historical business usage sub-data sequence set of the target push user set within a preset historical collection window; using K adaptive convolutional network layers to perform multi-scale feature extraction on the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set according to K receptive fields, respectively, to obtain K historical push feedback sub-data feature cluster sequence sets and K historical business usage sub-data feature sequence cluster sets, wherein the sizes of the K receptive fields are inconsistent; constructing an initial three-dimensional frame sequence, and combining the K historical push feedback sub-data feature cluster sequence sets and the K historical business usage sub-data The initial three-dimensional frame sequence is input according to the feature cluster sequence set to generate K historical push feedback frame sequence sets and K historical business use frame sequence sets; the K historical push feedback frame sequence sets and the K historical business use frame sequence sets are subjected to intra-sequence adjacency matrix fusion analysis to obtain K historical push feedback mapping feature sequence sets and K historical business use mapping feature sequence sets; the K historical push feedback mapping feature sequence sets and the K historical business use mapping feature sequence sets are transmitted to the fully connected layer for feature fusion analysis to obtain a user fusion feature sequence set; the user fusion feature sequence sets are reordered from large to small according to acceptance to obtain a user push fusion feature sequence set, and a user push time node sequence set is determined according to the user push fusion feature sequence set; the user personalized push scheme of the target push service is identified according to the user push time node sequence set to obtain a user personalized push scheme set, and SMS push management of the target push service is performed based on the user personalized push scheme set.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] By executing push feature collection, the push feature set of the target push service is obtained, ensuring the acquisition of basic data of the push service, including features such as service type, number of sub-services and sub-service type. These features provide a data basis for subsequent data analysis and push strategy optimization; the interactive intelligent robot obtains the historical behavior data of the target push user set of the target push service, including the historical push feedback sub-data sequence and the historical service usage sub-data sequence, providing comprehensive user behavior data, enabling the system to capture the user's push feedback and service usage in different time periods, providing basic data for personalized push; multi-scale feature extraction is performed using K adaptive convolutional network layers, Each convolutional layer extracts the features of historical data through different receptive fields, obtains feature cluster sequences of historical push feedback sub-data and feature sequence clusters of historical business usage sub-data of different scales, and captures features of different levels and scales in the data, so as to more comprehensively determine user behavior and business usage. This multi-scale feature extraction helps to improve the expressiveness and analysis accuracy of features. By constructing an initial three-dimensional frame sequence and inputting feature data into it, historical push feedback frame sequences and historical business usage frame sequences are generated. These frame sequences enable data to be organized and analyzed in the form of time series, which helps to identify temporal patterns and changing trends. The frame sequences are subjected to intra-sequence neighboring The mapping feature sequence is obtained by fusion analysis of the adjacency matrix. This process integrates feature information of different scales by constructing an adjacency matrix, performs deep feature interaction and fusion, enhances the interactivity and information fusion capabilities of the features, enables the model to better analyze the relationship between different features, and improves the effectiveness and accuracy of feature representation; the mapping feature sequence is transmitted to the fully connected layer for feature fusion analysis to obtain a set of user fusion feature sequences. The fully connected layer generates comprehensive behavioral features of users at different time nodes by comprehensively analyzing different features, and realizes deep fusion of features, so that the final user fusion feature sequence can fully reflect the user's behavior pattern and acceptance, thereby providing Personalized push provides more accurate data support; according to the acceptance ranking of the user's fusion feature sequence, the user push time node sequence set is determined. This step can identify the user's best push time point and formulate a suitable push timing for each user, optimizing the choice of push time so that the push content can be sent at the time when the user is most likely to accept it, thereby improving the push response rate and user satisfaction; personalized push schemes are identified according to the user push time node sequence set, and SMS push management is performed based on this, realizing personalized push management, so that the push content can more accurately meet the user's needs and preferences, thereby improving the push effect and user participation.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 A schematic diagram of the structure of an intelligent robot-assisted SMS push management system provided in an embodiment of the present application;

[0010] Figure 2 A schematic flow chart of the SMS push management method assisted by an intelligent robot provided in an embodiment of the present application.

[0011] Explanation of the accompanying drawings: push feature collection module 10, historical behavior data acquisition module 20, multi-scale feature extraction module 30, frame sequence acquisition module 40, adjacency matrix fusion analysis module 50, feature fusion analysis module 60, push time node determination module 70, SMS push management module 80. DETAILED DESCRIPTION

[0012] The embodiments of the present application provide an intelligent robot-assisted SMS push management system and method to solve the technical problem that traditional SMS push usually adopts a universal push strategy, fails to fully take into account the individual differences and behavior patterns of users, and cannot meet the individual needs of users, resulting in poor push effect. Through in-depth analysis of user behavior patterns, personalized push management is realized, thereby achieving the technical effect of improving push effect.

[0013] After introducing the basic principles of the present application, various non-limiting implementation methods of the present application will be specifically introduced below in conjunction with the accompanying drawings of the specification. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0014] Embodiment 1, as Figure 1 As shown, the embodiment of the present application provides an intelligent robot-assisted SMS push management system, the system comprising:

[0015] A push feature acquisition module 10, the push feature acquisition module 10 is used to perform push feature acquisition and obtain a push feature set of a target push service; a historical behavior data acquisition module 20, the historical behavior data acquisition module 20 is used for the interactive intelligent robot to obtain a set of historical behavior data sequences of a target push user set of the target push service collected in advance, wherein the set of historical behavior data sequences includes a set of historical push feedback sub-data sequences and a set of historical service usage sub-data sequences of the target push user set within a preset historical acquisition window; a multi-scale feature extraction module 30, the multi-scale feature extraction module 30 is used to respectively use K The adaptive convolutional network layer performs multi-scale feature extraction on the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set according to K receptive fields, respectively, to obtain K historical push feedback sub-data feature cluster sequence sets and K historical business usage sub-data feature sequence cluster sets, wherein the sizes of the K receptive fields are inconsistent; a frame sequence acquisition module 40, the frame sequence acquisition module 40 is used to construct an initial three-dimensional frame sequence, input the K historical push feedback sub-data feature cluster sequence sets and the K historical business usage sub-data feature cluster sequence sets into the initial three-dimensional frame sequence, and generate K historical push feedback frame sequences. The present invention relates to a plurality of historical business usage framework sequence sets and a plurality of historical business usage framework sequence sets; an adjacency matrix fusion analysis module 50, wherein the adjacency matrix fusion analysis module 50 is used to perform intra-sequence adjacency matrix fusion analysis on the K historical push feedback framework sequence sets and the K historical business usage framework sequence sets to obtain K historical push feedback mapping feature sequence sets and K historical business usage mapping feature sequence sets; a feature fusion analysis module 60, wherein the feature fusion analysis module 60 is used to transmit the K historical push feedback mapping feature sequence sets and the K historical business usage mapping feature sequence sets to the fully connected layer for feature fusion analysis to obtain a user fusion feature sequence. A column set; a push time node determination module 70, the push time node determination module 70 is used to reorder the user fusion feature sequence set according to the acceptance degree from large to small, obtain the user push fusion feature sequence set, and determine the user push time node sequence set according to the user push fusion feature sequence set; an SMS push management module 80, the SMS push management module 80 is used to identify the user personalized push plan of the target push service according to the user push time node sequence set, obtain the user personalized push plan set, and perform SMS push management of the target push service based on the user personalized push plan set.

[0016] Furthermore, the multi-scale feature extraction module 30 includes the following steps:

[0017] Perform information entropy analysis on the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set respectively to obtain a historical push information entropy set and a historical business usage information entropy set; perform centralized analysis on the historical push information entropy set and the historical business usage information entropy set to obtain K centralized information entropies; obtain multiple sample information entropies and multiple sample receptive fields as sample data to train a receptive field identifier until the training converges to obtain a trained receptive field identifier; send the K centralized information entropies to the receptive field identifier for analysis to obtain the K receptive fields.

[0018] Furthermore, the multi-scale feature extraction module 30 includes the following steps:

[0019] An information entropy distribution space is constructed for the historical push information entropy set and the historical business usage information entropy set, wherein the information entropy distribution space includes multiple spatial points, each spatial point corresponds to an information entropy; K spatial points are randomly selected from the multiple spatial points as K initial spatial points; the K initial spatial points are iteratively globally searched in the information entropy distribution space according to a preset iteration step size to obtain K iterative spatial point sets; the information entropy means of the K iterative spatial point sets are respectively calculated to obtain the K concentrated information entropies.

[0020] Furthermore, the framework sequence acquisition module 40 includes the following steps:

[0021] A central adaptive convolutional network layer is used to perform feature extraction on the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set according to a preset maximum receptive field to obtain a comprehensive feature cluster of the historical push feedback sub-data and a comprehensive feature cluster of the historical business usage sub-data; an initial three-dimensional framework is constructed based on the number of features in the comprehensive feature cluster of the historical push feedback sub-data and the comprehensive feature cluster of the historical business usage sub-data; the initial three-dimensional framework is copied and sorted according to the number of acquisition time nodes in a preset historical acquisition window to obtain the initial three-dimensional framework sequence.

[0022] Furthermore, the adjacency matrix fusion analysis module 50 includes the following operation steps:

[0023] A first historical push feedback frame sequence is randomly extracted from the K historical push feedback frame sequence sets; a first first historical push feedback frame in the first historical push feedback frame sequence is extracted as a first first historical push feedback mapping feature; a second first historical push feedback frame in the first historical push feedback frame sequence is extracted and an inner product calculation is performed with the first first historical push feedback frame to obtain a second adjacency matrix; a convolution operation is performed on the second adjacency matrix and the second first historical push feedback frame to obtain a second first historical push feedback mapping feature; a first historical push feedback mapping feature sequence is obtained after multiple adjacency matrix construction fusion analysis; an intra-sequence adjacency matrix fusion analysis is performed on the K historical push feedback frame sequence sets to obtain a K historical push feedback mapping feature sequence set.

[0024] Furthermore, the K receptive fields are the feature extraction widths of the K adaptive convolutional network layers on the input data.

[0025] Furthermore, the user fusion feature sequence is a sequence formed by the acceptance of the target push service by users in the target push user set at different time points.

[0026] Through the subsequent detailed description of the intelligent robot-assisted SMS push management method in this specification, technical personnel in this field can clearly understand the intelligent robot-assisted SMS push management system in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0027] Embodiment 2 is based on the same inventive concept as the intelligent robot-assisted SMS push management system in the previous embodiment. Figure 2 As shown, the embodiment of the present application provides an intelligent robot-assisted SMS push management method, the method comprising:

[0028] Perform push feature collection to obtain the push feature set of the target push service.

[0029] First, clarify the target push business and determine the type of push business, such as marketing promotion, service update, notification reminder, etc. The push business contains multiple sub-businesses. For example, in a marketing campaign, the sub-business may be a different product or service promotion. Determine the number of sub-businesses. For example, a push campaign contains multiple sub-businesses, and each sub-business may be pushed to different user groups or at different time points. According to the target push business, extract relevant business feature information from different data sources, including but not limited to business type, number of sub-businesses, sub-business type, target user group characteristics, historical push effects, etc., and integrate the extracted features to form a push feature set.

[0030] The interactive intelligent robot obtains a pre-collected historical behavior data sequence set of the target push user set of the target push service, wherein the historical behavior data sequence set includes a historical push feedback sub-data sequence set and a historical service usage sub-data sequence set of the target push user set within a preset historical collection window.

[0031] The intelligent robot is a data extraction tool. Its main task is to automatically retrieve and extract the corresponding historical behavior data based on the target push user set. First, it is necessary to clarify the user set targeted by the target push service, that is, the target push user set. These users may be groups that have performed certain behaviors or participated in similar push services, such as users who have purchased specific products or subscribed to related services.

[0032] The preset historical collection window is a preset time range that can be set according to business needs or push frequency. The user's historical behavior data is collected within this time range. For example, a 24-hour collection window can be defined to collect the user's historical push feedback sub-data and historical business usage sub-data every hour within this 24 hours. Among them, the historical push feedback sub-data refers to the historical push feedback data of the target push user, for example, a series of feedback behaviors that the user may generate after receiving the previous push message, such as click, ignore, reply, etc.; the historical behavior data is the target push user's usage data of related services, for example, the user's frequency of use of a certain service over a period of time, duration, function preference, etc., these data form the business usage sub-data.

[0033] The extracted historical behavior data is serialized, that is, the original discrete data points are organized into data sequences in chronological order, and then the data sequences of all users are integrated to obtain a set of historical push feedback sub-data sequences and a set of historical business usage sub-data sequences. These sequences reflect the temporal evolution of user behavior and provide a basis for subsequent multi-scale feature extraction.

[0034] K adaptive convolutional network layers are used to perform multi-scale feature extraction on the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set according to K receptive fields, to obtain K historical push feedback sub-data feature cluster sequence sets and K historical business usage sub-data feature sequence cluster sets, wherein the sizes of the K receptive fields are inconsistent.

[0035] An adaptive convolutional network layer refers to a convolutional layer that can dynamically adjust its own parameters according to the input data during the feature extraction process. This adaptability enables the network to better capture the key features in the input data.

[0036] The receptive field refers to the range of input data that a neuron in a certain layer of the network can capture. In convolutional neural networks, the size of the receptive field is determined by factors such as the size of the convolution kernel, the step size, and the parameters of the pooling layer. The size of the receptive field directly affects the network's depth of understanding of the data and its ability to capture details. Using K receptive fields of different sizes means that each convolution layer uses a different convolution kernel size and step size to process data. Some layers pay more attention to local details, while others pay more attention to overall trends.

[0037] The historical push feedback sub-data sequence set and the historical business usage sub-data sequence set are used as input data and input into K adaptive convolutional network layers. K receptive fields are used for multi-scale feature extraction, that is, each convolutional layer uses different convolution kernel sizes and step sizes to process data. This setting allows the network to capture features at different scales. For example, a small receptive field can capture tiny details in the data, while a large receptive field can extract global patterns. This inconsistency enables the network to understand the input data from multiple perspectives, thereby extracting richer features.

[0038] After being processed by K convolutional layers with different receptive fields, the historical push feedback sub-data is refined into K different feature cluster sequence sets. Each set corresponds to a receptive field and represents the features extracted at that scale. Similarly, the historical service usage sub-data is also processed by K convolutional layers to generate K different feature cluster sequence sets. These sets reflect the service usage patterns of users in different time periods and usage scenarios.

[0039] An initial three-dimensional frame sequence is constructed, and the K historical push feedback sub-data feature cluster sequence sets and the K historical business usage sub-data feature cluster sequence sets are input into the initial three-dimensional frame sequence to generate K historical push feedback frame sequence sets and K historical business usage frame sequence sets.

[0040] Construct an initial three-dimensional frame sequence. The three-dimensional frame is a multi-dimensional structure. It not only studies the feature data itself, but also introduces the time dimension and the relationship dimension between different features. Through the three-dimensional frame, these feature sequences can be represented and organized in a unified space, so as to better capture the temporal patterns and relationships in the data.

[0041] The K historical push feedback sub-data feature cluster sequence sets and the K historical business usage sub-data feature cluster sequence sets are mapped to the corresponding positions of the initial three-dimensional frame respectively. Each feature cluster sequence corresponds to a position in the three-dimensional frame and is arranged in chronological order. The frame diagram is a slice in the three-dimensional frame sequence, which corresponds to a specific time point or feature type. Through the frame diagram, the changes in user behavior or business usage characteristics at a specific time point can be intuitively displayed. Each frame diagram corresponds to a feature sequence, thereby constructing a time series connection for the user's behavior.

[0042] By arranging K historical push feedback sub-data feature cluster sequences according to time and feature dimensions, K historical push feedback frame sequence sets are generated. Each frame sequence is a collection of a set of frame diagrams. These frame diagrams are connected by time series and reflect the user's behavior patterns in historical push. Similarly, by mapping K historical business usage sub-data feature cluster sequences into a three-dimensional framework, K historical business usage frame sequence sets are generated. These sequences show the user's business usage in different time periods and the evolution of these usages.

[0043] An intra-sequence adjacency matrix fusion analysis is performed on the K historical push feedback framework sequence sets and the K historical service usage framework sequence sets to obtain K historical push feedback mapping feature sequence sets and K historical service usage mapping feature sequence sets.

[0044] In graph theory, the adjacency matrix is ​​used to represent the connection relationship between nodes in a graph. In this step, the adjacency matrix is ​​used to represent the temporal relationship between each feature point in a feature sequence. By constructing an intra-sequence adjacency matrix, the dependency and association patterns within the feature sequence can be captured. That is, if two feature points are adjacent in the time series, then a connection can be established between them.

[0045] By performing fusion analysis on the adjacency matrix within each frame sequence, the feature points with strong correlation in the same sequence can be integrated to generate correlated feature representations, which can better capture the potential patterns of user behavior and business usage characteristics. By performing adjacency matrix fusion analysis on the K historical push feedback frame sequence sets, K historical push feedback mapping feature sequence sets are generated. These mapping feature sequences are further abstractions of the original features and contain temporal correlations. Similarly, by performing adjacency matrix fusion analysis on the K historical business usage frame sequence sets, K historical business usage mapping feature sequence sets are generated. These feature sequences can better reflect the user's behavior patterns in different business usage scenarios.

[0046] The K historical push feedback mapping feature sequence sets and the K historical business usage mapping feature sequence sets are transmitted to the fully connected layer for feature fusion analysis to obtain a user fusion feature sequence set.

[0047] The fully connected layer is a basic layer in the neural network, which is used to combine different features into a higher-level representation. In the fully connected layer, each input node is connected to the output node, so that the input features can generate new output features through weighted or nonlinear activation functions. Feature fusion refers to the merging of features from different sources or different types to generate a unified feature representation.

[0048] The K historical push feedback mapping feature sequence sets and the K historical business usage mapping feature sequence sets are used as input and transmitted to the fully connected layer. In the fully connected layer, the weight matrix is ​​used to weight the input feature sequence. Each input feature point will be multiplied by the corresponding weight value to generate a weighted feature value. On the basis of the weighted feature value, a bias value is added to further adjust the feature output. Furthermore, in the fully connected layer, nonlinear activation functions such as ReLU are used to perform nonlinear mapping on the weighted and biased features, aiming to introduce nonlinear characteristics and enhance the model's ability to capture complex feature relationships. The connection layer generates a new feature representation by weighting, biasing and activating the input feature sequence, and outputs a user fusion feature sequence set. This set represents the user's behavioral characteristics in multiple dimensions and is a comprehensive feature sequence set. The user fusion feature sequence is a sequence formed by the acceptance of the target push service by users in the target push user set at different time nodes.

[0049] The user fusion feature sequence set is reordered from large to small according to acceptance degree to obtain a user push fusion feature sequence set, and a user push time node sequence set is determined according to the user push fusion feature sequence set.

[0050] In the push scenario, acceptance is an indicator to measure the user's preference for push content or timing. Acceptance can be calculated based on historical data, behavior patterns, user feedback and other factors. For each user's fused feature sequence, the corresponding acceptance score is calculated. This score can be obtained based on a weighted combination of multiple factors, such as the click-through rate, reading time, interaction frequency in the user's historical behavior. According to the calculated acceptance score, the user fused feature sequence set is re-sorted in descending order to obtain the user push fused feature sequence set. The sorted sequence reflects the user's acceptance at different time points.

[0051] According to the high acceptance time period in the user push fusion feature sequence set, select the appropriate push time node, that is, select the time point with higher acceptance for push, so as to ensure that the push content is delivered to the user at the best time, and combine the selected push time nodes into a user push time node sequence set as a reference for the final push strategy. This time node sequence set guides the system to push at the appropriate time, thereby improving the user's acceptance rate and interaction rate. For example, if users interact frequently in certain specific time periods, these time periods can be preferentially selected as push time nodes.

[0052] The user personalized push scheme of the target push service is identified according to the user push time node sequence set, a user personalized push scheme set is obtained, and SMS push management of the target push service is performed based on the user personalized push scheme set.

[0053] Based on the user's historical behavior and preferences, we identify the push content that best meets their needs, match the push content with the time node, and form a personalized push plan for the user. For example, if a user's acceptance rate is highest at 9 pm, then we will choose to push relevant content to him during this time period. Based on each user's push content and the best time node, we generate a set of personalized push plans, each of which covers push content, push time, and push method.

[0054] Based on the user's personalized push plan set, the optimal push time and content are selected and sent to the user through the SMS channel. Through precise push plans and optimized push timing, the user experience can be significantly improved, interruptions can be reduced, and user acceptance of push content can be increased.

[0055] Further, including:

[0056] Perform information entropy analysis on the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set respectively to obtain a historical push information entropy set and a historical business usage information entropy set; perform centralized analysis on the historical push information entropy set and the historical business usage information entropy set to obtain K centralized information entropies; obtain multiple sample information entropies and multiple sample receptive fields as sample data to train a receptive field identifier until the training converges to obtain a trained receptive field identifier; send the K centralized information entropies to the receptive field identifier for analysis to obtain the K receptive fields.

[0057] Information entropy is an indicator that measures the uncertainty of random variables. The higher the information entropy, the greater the uncertainty and the richer the information. In push and business usage data, information entropy is used to identify those parts with higher uncertainty and important information.

[0058] First, the historical push feedback sub-data series set and the historical business usage sub-data series set are preprocessed, such as denoising and normalization, to ensure that the data is suitable for information entropy calculation. The information entropy of each data series is calculated. The common calculation method is to count the probability of occurrence of each possible value, and then use the information entropy formula to calculate the entropy value. Finally, the historical push information entropy set and the historical business usage information entropy set are generated, which reflect the amount of information and uncertainty in the historical push feedback data and business usage data, respectively.

[0059] The concentrated information entropy is an indicator obtained by aggregate analysis of multiple information entropy sets, which is used to identify the main sources of uncertainty. The historical push information entropy set and the historical business use information entropy set are aggregated to construct an information entropy distribution space, and an iterative global search of K initial space points is performed based on the information entropy distribution space. The information entropy mean is calculated based on the global search results to obtain the K concentrated information entropies.

[0060] The receptive field identifier is a machine learning model used to identify the characteristics of the receptive field. In a convolutional neural network, the receptive field refers to the size of the input area perceived by each neuron in the network. By training the receptive field identifier, we can determine which receptive fields are most effective for extracting data features.

[0061] Multiple sample information entropies are calculated from historical data to represent the distribution of information and uncertainty in the data; multiple sample receptive fields are extracted from different convolutional network models to represent the perception range of different receptive fields on the data. Multiple sample information entropies and multiple sample receptive fields are used as training data sets, and these data are standardized to facilitate model training. A receptive field identifier model is designed based on a neural network, and sample information entropy and sample receptive field are used for training. During the training process, the model learns how to identify and determine the most effective receptive field based on information entropy features. The training error and verification error are monitored to determine whether the model has been trained to convergence, for example, the error no longer decreases significantly. When the training converges, an optimized receptive field identifier is obtained, which can accurately identify important receptive field features in the data.

[0062] The K concentrated information entropies are input into the trained receptive field identifier. The receptive field identifier analyzes the input data and uses its learned model parameters to identify and determine the most effective receptive field. Finally, K receptive fields are output, which represent the most suitable receptive field features under different concentrated information entropies, and are used to extract key features in the data.

[0063] Further, including:

[0064] An information entropy distribution space is constructed for the historical push information entropy set and the historical business usage information entropy set, wherein the information entropy distribution space includes multiple spatial points, each spatial point corresponds to an information entropy; K spatial points are randomly selected from the multiple spatial points as K initial spatial points; the K initial spatial points are iteratively globally searched in the information entropy distribution space according to a preset iteration step size to obtain K iterative spatial point sets; the information entropy means of the K iterative spatial point sets are respectively calculated to obtain the K concentrated information entropies.

[0065] The information entropy distribution space is a space representing the distribution of information entropy values. In this distribution space, a spatial point is a specific instance of an information entropy value. Each point represents a data point in a historical push information entropy set or a historical business use information entropy set, reflecting the amount of information and uncertainty of the point in the data. Specifically, information entropy values ​​are extracted from the historical push information entropy set and the historical business use information entropy set, and the extracted information entropy values ​​are mapped to a multidimensional space. Each information entropy value is regarded as a point in the space, and the entire space is composed of these points, which is the information entropy distribution space.

[0066] K spatial points are randomly selected from multiple spatial points in the constructed distribution space. The purpose of random selection is to ensure that the initial points cover different areas in the space, thereby providing diverse initial conditions. The randomly selected K spatial points are used as K initial spatial points, which are the points used to initialize the algorithm.

[0067] The preset iteration step size is the adjustment amplitude of each step in the iteration process, which determines the movement range of the spatial point in the distribution space. The choice of step size affects the efficiency and effect of the search. Iterative global search refers to iterating the initial spatial point in the information entropy distribution space to find the spatial point with the best characteristics, and continuously adjusts the position of the spatial point to achieve the global optimal or near-optimal state in the space.

[0068] Specifically, K randomly selected initial spatial points are used as starting points. For each initial spatial point, the information entropy difference between it and the surrounding spatial points is calculated, and the spatial points whose difference is within a preset threshold are added to the iterative spatial point set. These points can provide more valuable information. The position of the spatial point is adjusted according to the preset iteration step to explore different areas of the information entropy distribution space. The above steps are repeated until the iteration stop condition is met, such as reaching the maximum number of iterations or the change in the information entropy difference is less than the preset threshold. At this time, K iterative spatial point sets are output.

[0069] The information entropy value of each spatial point is accumulated, and the total information entropy value is divided by the number of spatial points to obtain the concentrated information entropy. The mean of each iterative spatial point set is calculated in turn to obtain the final K concentrated information entropies, which represent the overall information entropy characteristics of these points in the information entropy distribution space.

[0070] Further, including:

[0071] A central adaptive convolutional network layer is used to perform feature extraction on the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set according to a preset maximum receptive field to obtain a comprehensive feature cluster of the historical push feedback sub-data and a comprehensive feature cluster of the historical business usage sub-data; an initial three-dimensional framework is constructed based on the number of features in the comprehensive feature cluster of the historical push feedback sub-data and the comprehensive feature cluster of the historical business usage sub-data; the initial three-dimensional framework is copied and sorted according to the number of acquisition time nodes in a preset historical acquisition window to obtain the initial three-dimensional framework sequence.

[0072] The preset maximum receptive field refers to the maximum range of the convolution kernel in the convolution network layer. A larger receptive field can capture features in a wider range. The central adaptive convolution network layer is used to extract features from the historical push feedback sub-data sequence set and the historical business usage sub-data sequence set according to the preset maximum receptive field, that is, the convolution kernel is used to slide on the input data, and the convolution operation is performed to extract features. According to the characteristics of the input data, the size of the convolution kernel will be dynamically adjusted to adapt to the characteristics of different data. After the convolution operation, the feature map is output to represent the features extracted under different receptive fields. Through the convolution operation, a comprehensive feature cluster of the historical push feedback sub-data and a comprehensive feature cluster of the historical business usage sub-data are obtained, which comprehensively represent various features in the historical push feedback data and various features in the historical business usage data.

[0073] A 3D frame is a spatial representation that maps feature data into a 3D coordinate system. This frame is used to visualize different features in the data and the relationships between them. Count the number of all features in the comprehensive feature cluster of historical push feedback sub-data and the comprehensive feature cluster of historical business usage sub-data, draw a 3D frame in 3D space, map the number of features into the 3D frame, and build an initial 3D frame, so that the relationship and distribution of features can be analyzed more intuitively.

[0074] According to the number of collection time nodes in the preset historical collection window, the initial three-dimensional frame is copied. Each copied frame represents the data at a time point. The copied three-dimensional frames are sorted in chronological order to form a time series, in which each time node is assigned a three-dimensional frame to generate an initial three-dimensional frame sequence. This sequence shows the changes in historical push feedback and business usage data over time.

[0075] Further, including:

[0076] A first historical push feedback frame sequence is randomly extracted from the K historical push feedback frame sequence sets; a first first historical push feedback frame in the first historical push feedback frame sequence is extracted as a first first historical push feedback mapping feature; a second first historical push feedback frame in the first historical push feedback frame sequence is extracted and an inner product calculation is performed with the first first historical push feedback frame to obtain a second adjacency matrix; a convolution operation is performed on the second adjacency matrix and the second first historical push feedback frame to obtain a second first historical push feedback mapping feature; a first historical push feedback mapping feature sequence is obtained after multiple adjacency matrix construction fusion analysis; an intra-sequence adjacency matrix fusion analysis is performed on the K historical push feedback frame sequence sets to obtain a K historical push feedback mapping feature sequence set.

[0077] A frame sequence is randomly selected from the set of K historical push feedback frame sequences as an analysis object, which is the first historical push feedback frame sequence.

[0078] The first frame is extracted from the first historical push feedback frame sequence. The first frame in the sequence refers to the frame of the earliest data point in the time series, indicating the state at the beginning of the sequence. The first frame is used as the first first historical push feedback mapping feature.

[0079] A second frame is extracted from the first historical push feedback frame sequence, where the second frame in the sequence refers to a frame that follows the first frame in the time series.

[0080] The inner product is a standard method for calculating the similarity between two vectors. The inner product is calculated for the second first history push feedback frame and the first first history push feedback frame to obtain the similarity between them. According to the result of the inner product calculation, the second adjacency matrix is ​​generated. The adjacency matrix is ​​a matrix that represents the relationship between data points. In this step, the adjacency matrix is ​​used to represent the association relationship between the first frame and the second frame.

[0081] The convolution operation is the process of performing weighted summation on the input data by sliding the convolution kernel. The convolution kernel size and weight are set according to requirements, and the convolution kernel is slid on the second adjacency matrix and the second frame. The weighted sum is calculated to obtain the convolution result, which is used as the second first history push feedback mapping feature, indicating the feature extracted in the convolution operation.

[0082] In this way, adjacency matrices are constructed for multiple historical push feedback frame sequences in turn, and these matrices are fused and analyzed to obtain a more comprehensive feature sequence, thereby obtaining the first historical push feedback mapping feature sequence.

[0083] In this way, the K historical push feedback frame sequence sets are sequentially subjected to intra-sequence adjacency matrix fusion analysis to obtain K historical push feedback mapping feature sequence sets.

[0084] Furthermore, the K receptive fields are the feature extraction widths of the K adaptive convolutional network layers on the input data.

[0085] K receptive fields means that in K adaptive convolutional networks, K different receptive field sizes are used to extract features. These receptive fields can adapt to different input data features to capture local information of different scales. Each adaptive convolutional network layer has a different receptive field width, which enables the network to extract local features of data from different scales. For example, one layer uses a 3x3 convolution kernel to extract local details, while another layer uses a 7x7 convolution kernel to extract larger regional features. By fusing features from different receptive fields, a multi-scale understanding of the data can be obtained for processing tasks with different spatial scales.

[0086] Furthermore, the user fusion feature sequence is a sequence formed by the acceptance of the target push service by users in the target push user set at different time points.

[0087] The user fusion feature sequence is a record of changes in each user's acceptance of the target push service. At different time points, the user's behavior pattern is determined based on the fusion feature sequence, such as the user's click rate, response time, and interaction frequency. Through the fusion analysis of the fully connected layer, the fusion features of each user at each time node are obtained to form a time series. These sequences reflect the changes in the user's acceptance of the pushed content and help the system identify the user's behavior trends and preferences.

[0088] In summary, the intelligent robot-assisted SMS push management method provided in the embodiments of the present application has the following technical effects:

[0089] By executing push feature collection, the push feature set of the target push service is obtained, ensuring the acquisition of basic data of the push service, including features such as service type, number of sub-services and sub-service type. These features provide a data basis for subsequent data analysis and push strategy optimization; the interactive intelligent robot obtains the historical behavior data of the target push user set of the target push service, including the historical push feedback sub-data sequence and the historical service usage sub-data sequence, providing comprehensive user behavior data, enabling the system to capture the user's push feedback and service usage in different time periods, providing basic data for personalized push; multi-scale feature extraction is performed using K adaptive convolutional network layers, Each convolutional layer extracts the features of historical data through different receptive fields, obtains feature cluster sequences of historical push feedback sub-data and feature sequence clusters of historical business usage sub-data of different scales, and captures features of different levels and scales in the data, so as to more comprehensively determine user behavior and business usage. This multi-scale feature extraction helps to improve the expressiveness and analysis accuracy of features. By constructing an initial three-dimensional frame sequence and inputting feature data into it, historical push feedback frame sequences and historical business usage frame sequences are generated. These frame sequences enable data to be organized and analyzed in the form of time series, which helps to identify temporal patterns and changing trends. The frame sequences are subjected to intra-sequence neighboring The mapping feature sequence is obtained by fusion analysis of the adjacency matrix. This process integrates feature information of different scales by constructing an adjacency matrix, performs deep feature interaction and fusion, enhances the interactivity and information fusion capabilities of the features, enables the model to better analyze the relationship between different features, and improves the effectiveness and accuracy of feature representation; the mapping feature sequence is transmitted to the fully connected layer for feature fusion analysis to obtain a set of user fusion feature sequences. The fully connected layer generates comprehensive behavioral features of users at different time nodes by comprehensively analyzing different features, and realizes deep fusion of features, so that the final user fusion feature sequence can fully reflect the user's behavior pattern and acceptance, thereby providing Personalized push provides more accurate data support; according to the acceptance ranking of the user's fusion feature sequence, the user push time node sequence set is determined. This step can identify the user's best push time point and formulate a suitable push timing for each user, optimizing the choice of push time so that the push content can be sent at the time when the user is most likely to accept it, thereby improving the push response rate and user satisfaction; personalized push schemes are identified according to the user push time node sequence set, and SMS push management is performed based on this, realizing personalized push management, so that the push content can more accurately meet the user's needs and preferences, thereby improving the push effect and user participation.

[0090] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. Intelligent robot-assisted SMS push management system, It is characterized in that The system comprises: A push feature collection module, which is used to perform push feature collection and obtain a push feature set of a target push service; A historical behavior data acquisition module, wherein the historical behavior data acquisition module is used for the interactive intelligent robot to acquire a pre-collected historical behavior data sequence set of a target push user set of the target push service, wherein the historical behavior data sequence set includes a historical push feedback sub-data sequence set and a historical service usage sub-data sequence set of the target push user set within a preset historical collection window; A multi-scale feature extraction module, the multi-scale feature extraction module is used to use K adaptive convolutional network layers to perform multi-scale feature extraction on the historical push feedback sub-data sequence set and the historical service usage sub-data sequence set according to K receptive fields, to obtain K historical push feedback sub-data feature cluster sequence sets and K historical service usage sub-data feature cluster sequence sets, wherein the sizes of the K receptive fields are inconsistent; A frame sequence acquisition module, wherein the frame sequence acquisition module is used to construct an initial three-dimensional frame sequence, input the K historical push feedback sub-data feature cluster sequence sets and the K historical business use sub-data feature cluster sequence sets into the initial three-dimensional frame sequence, and generate K historical push feedback frame sequence sets and K historical business use frame sequence sets; An adjacency matrix fusion analysis module, which is used to perform intra-sequence adjacency matrix fusion analysis on the K historical push feedback framework sequence sets and the K historical business usage framework sequence sets to obtain K historical push feedback mapping feature sequence sets and K historical business usage mapping feature sequence sets; A feature fusion analysis module, wherein the feature fusion analysis module is used to transmit the K historical push feedback mapping feature sequence sets and the K historical business usage mapping feature sequence sets to the fully connected layer for feature fusion analysis to obtain a user fusion feature sequence set; A push time node determination module, the push time node determination module is used to reorder the user fusion feature sequence set from large to small according to acceptance, obtain a user push fusion feature sequence set, and determine a user push time node sequence set according to the user push fusion feature sequence set; An SMS push management module, the SMS push management module is used to identify the user personalized push scheme of the target push service according to the user push time node sequence set, obtain the user personalized push scheme set, and perform SMS push management of the target push service based on the user personalized push scheme set; The adjacency matrix fusion analysis module is further used to perform the following steps: Randomly extracting a first historical push feedback frame sequence from the K historical push feedback frame sequence sets; extracting a first first historical push feedback frame in the first historical push feedback frame sequence as a first first historical push feedback mapping feature; Extracting a second first history push feedback frame in the first history push feedback frame sequence and performing inner product calculation on the first first history push feedback frame to obtain a second adjacency matrix; Performing a convolution operation on the second adjacency matrix and the second first history push feedback framework to obtain a second first history push feedback mapping feature; After multiple adjacency matrix construction and fusion analysis, the first historical push feedback mapping feature sequence is obtained; An intra-sequence adjacency matrix fusion analysis is performed on the K historical push feedback framework sequence sets to obtain K historical push feedback mapping feature sequence sets.

2. The intelligent robot-assisted SMS push management system according to claim 1, It is characterized in that The multi-scale feature extraction module includes the following steps: Performing information entropy analysis on the historical push feedback sub-data sequence set and the historical service usage sub-data sequence set respectively to obtain a historical push information entropy set and a historical service usage information entropy set; Centrally analyzing the historical push information entropy set and the historical service usage information entropy set to obtain K centralized information entropies; Obtain multiple sample information entropies and multiple sample receptive fields as sample data to train a receptive field identifier until the training converges, thereby obtaining a trained receptive field identifier; The K concentrated information entropies are sent to the receptive field identifier for analysis to obtain the K receptive fields.

3. The intelligent robot-assisted SMS push management system as claimed in claim 2, It is characterized in that The multi-scale feature extraction module includes the following steps: Constructing an information entropy distribution space for the historical push information entropy set and the historical business usage information entropy set, wherein the information entropy distribution space includes a plurality of space points, each space point corresponding to an information entropy; Randomly select K spatial points from the multiple spatial points as K initial spatial points; Iteratively perform global search on the K initial spatial points in the information entropy distribution space according to a preset iteration step size to obtain K iterative spatial point sets; The information entropy means of the K iterative spatial point sets are calculated respectively to obtain the K concentrated information entropies.

4. The intelligent robot-assisted SMS push management system as claimed in claim 3, It is characterized in that The framework sequence acquisition module includes the following steps: Using a central adaptive convolutional network layer, feature extraction is performed on the historical push feedback sub-data sequence set and the historical service usage sub-data sequence set according to a preset maximum receptive field to obtain a comprehensive feature cluster of the historical push feedback sub-data and a comprehensive feature cluster of the historical service usage sub-data; Constructing an initial three-dimensional framework based on the number of features in the comprehensive feature cluster of the historical push feedback sub-data and the comprehensive feature cluster of the historical business usage sub-data respectively; The initial three-dimensional frame is copied and sorted according to the number of acquisition time nodes in a preset historical acquisition window to obtain the initial three-dimensional frame sequence.

5. The intelligent robot-assisted SMS push management system according to claim 1, It is characterized in that The K receptive fields are the feature extraction widths of the K adaptive convolutional network layers on the input data.

6. The intelligent robot-assisted SMS push management system according to claim 1, It is characterized in that in, The user fusion feature sequence is a sequence formed by the acceptance of the target push service by users in the target push user set at different time points.

7. Intelligent robot-assisted SMS push management method, It is characterized in that Based on the implementation of the intelligent robot-assisted SMS push management system according to any one of claims 1 to 6, the method comprises: Execute push feature collection to obtain the push feature set of the target push service; The interactive intelligent robot obtains a previously collected historical behavior data sequence set of a target push user set of the target push service, wherein the historical behavior data sequence set includes a historical push feedback sub-data sequence set and a historical service usage sub-data sequence set of the target push user set within a preset historical collection window; Using K adaptive convolutional network layers respectively, perform multi-scale feature extraction on the historical push feedback sub-data sequence set and the historical service usage sub-data sequence set according to K receptive fields to obtain K historical push feedback sub-data feature cluster sequence sets and K historical service usage sub-data feature cluster sequence sets, wherein the sizes of the K receptive fields are inconsistent; Constructing an initial three-dimensional frame sequence, inputting the K historical push feedback sub-data feature cluster sequence sets and the K historical business usage sub-data feature cluster sequence sets into the initial three-dimensional frame sequence, generating K historical push feedback frame sequence sets and K historical business usage frame sequence sets; Performing intra-sequence adjacency matrix fusion analysis on the K historical push feedback framework sequence sets and the K historical service usage framework sequence sets to obtain K historical push feedback mapping feature sequence sets and K historical service usage mapping feature sequence sets; The K historical push feedback mapping feature sequence sets and the K historical business usage mapping feature sequence sets are transmitted to the fully connected layer for feature fusion analysis to obtain a user fusion feature sequence set; Reorder the user fusion feature sequence set from large to small according to acceptance degree to obtain a user push fusion feature sequence set, and determine a user push time node sequence set according to the user push fusion feature sequence set; The user personalized push scheme of the target push service is identified according to the user push time node sequence set, a user personalized push scheme set is obtained, and SMS push management of the target push service is performed based on the user personalized push scheme set.

Citation Information

Patent Citations

  • Information pushing method and device, storage medium and computer equipment

    CN110825968A

  • Information recommendation method and device based on artificial intelligence, electronic equipment and storage medium

    CN111444428A