A network information propagation method and system

Abstract

The application relates to the technical field of network communication and discloses a network information propagation method and system, wherein an information relay node maintains a community characteristic matrix for each community to dynamically represent a community information state; after receiving to-be-propagated information, a feature vector of the to-be-propagated information is extracted and similarity calculation and fusion are performed with a current state of the matrix, and then information entropy variation before and after matrix updating is calculated to quantize a propagation value weight of the information to the community; a node combines real-time community space-time context features to predict expected coverage utility of different forwarding strategies; finally, actual propagation effects and resident feedback are collected, difference analysis is performed on the actual propagation effects and the resident feedback and the prediction results, and the prediction model and the community characteristic matrix are updated online according to the difference analysis, so that accurate quantization of information value, intelligent decision-making of scene self-adaptation and continuous self-optimization of the system are realized, and the efficiency of community information propagation and resource utilization are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network communication, and discloses a network information propagation method and system. BACKGROUND

[0002] With the rapid development of mobile Internet and Internet of Things technology, community networks have become an important carrier of information dissemination. In the scenarios of smart communities, emergency communication, and local service push, how to efficiently and accurately disseminate information to the target group while avoiding network congestion and resource waste is a key challenge.

[0003] Existing network information propagation methods mostly use topology-based flooding, probability-based forwarding, or simple history-based filtering mechanisms. These methods still have many shortcomings, such as lack of quantitative evaluation of the value of information content itself, relying solely on network structure or fixed probability for forwarding, which can easily lead to insufficient dissemination of important information and excessive dissemination of unimportant information; the decision-making process does not fully consider the dynamic spatiotemporal properties of community-specific social spaces, such as resident activity patterns and community schedules, resulting in inaccurate dissemination timing and path selection; existing methods are mostly open-loop systems that cannot dynamically optimize based on actual information dissemination results and real feedback from residents, resulting in poor adaptability. SUMMARY

[0004] This section aims to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, abstract, and title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0005] To solve the above technical problems, the present application provides a network information propagation method and system.

[0006] In one aspect, the present application provides a network information propagation method, comprising S1, an information relay node sets a community feature matrix for each community, when receiving a to-be-propagated information, extracts a feature vector of the to-be-propagated information, calculates the similarity of the feature vector and the column vector of the community feature matrix of the target community, integrates the feature vector into the current time column of the community feature matrix, and obtains an updated community feature matrix;

[0007] S2, calculating the entropy value of the updated community feature matrix, taking the change of the entropy value as the propagation value weight of the to-be-propagated information for the community;

[0008] S3, the information relay node obtains the community spatiotemporal context feature at the current time, inputs the propagation value weight and the community spatiotemporal context feature into a forwarding utility model, and predicts the coverage utility generated by forwarding information by each neighbor node in the target community.

[0009] S4, collecting the feedback matrix constructed by the resident feedback, the resident demand feedback vector and the actual coverage utility matrix, calculating the difference between the actual coverage utility matrix and the expected coverage utility matrix, and updating the forwarding utility model and the community feature matrix combined with the feedback matrix.

[0010] As a preferred scheme of the network information propagation method of the present application, wherein:

[0011] The behavior information feature dimension of the community feature matrix is columned as the time dimension, and the matrix element represents the probability of the feature appearing in the corresponding time window;

[0012] The community feature matrix is set for each community, including defining the row of the community feature matrix as the information feature dimension, which is used to represent different types of information content attributes;

[0013] The column of the community feature matrix is defined as the time window, which is used to represent the time sequence evolution of the community information state;

[0014] Each element value in the community feature matrix is the probability estimate of the corresponding feature appearing or propagating in the community in the corresponding time window.

[0015] As a preferred scheme of the network information propagation method of the present application, wherein:

[0016] By keyword extraction, topic classification and embedding representation on the to-be-propagated information, a normalized feature vector aligned with the row dimension of the community feature matrix is generated;

[0017] From the community feature matrix, the column vector of the current time window is selected as the target vector;

[0018] The cosine similarity between the information feature vector and the target column vector is calculated as the similarity measure.

[0019] As a preferred scheme of the network information propagation method of the present application, wherein:

[0020] The similarity calculation fusion weight is used to interpolate the information feature vector and the current time column vector of the community feature matrix, and the updated current time column vector is obtained;

[0021] The updated column vector replaces the corresponding column in the original community feature matrix, and the community feature matrix is updated.

[0022] As a preferred scheme of the network information propagation method of the present application, wherein:

[0023] Each column of the community feature matrix is ​​regarded as a discrete probability distribution vector describing the distribution of community information features within a time window;

[0024] For the community feature matrices before and after the update, calculate the Shannon entropy of each probability distribution vector column by column to obtain the entropy value sequence arranged by time window;

[0025] The entropy values ​​of each time window are weighted and summed according to time weights to obtain the weighted total entropy of the community feature matrix before and after the update.

[0026] The change in entropy is calculated as the difference between the weighted total entropy before the update and the weighted total entropy after the update. The difference is used as the information gain that the information to be disseminated brings to the target community. The greater the entropy reduction, the higher the value. The value is used directly or after normalization as the dissemination value weight of the information to be disseminated to the community.

[0027] As a preferred embodiment of the network information dissemination method of this application, wherein:

[0028] The propagation value weights and community spatiotemporal context features are input into the forwarding utility model, including:

[0029] The information relay node obtains its own real-time geographical location and timestamp through a built-in positioning and timing module; and obtains community affiliation identifier, real-time location, movement speed and direction by exchanging signaling with neighboring nodes.

[0030] The feature vector formed by the propagation value weight and the acquired context features is input into the forwarding utility model.

[0031] As a preferred embodiment of the network information dissemination method of this application, wherein:

[0032] The forwarding utility model takes the fused feature vector as input, forwards information to different neighbor nodes of the information relay point, and predicts the expected information coverage increment and propagation efficiency gain in the target community within a preset time window in the future, with the expected information coverage increment and the propagation efficiency gain as the output targets.

[0033] The initial training of the forwarding utility model is based on a historical propagation record dataset, which contains feature inputs and actual propagation effect sample pairs.

[0034] The forwarding utility model is established by learning the nonlinear relationships in historical data.

[0035] As a preferred embodiment of the network information dissemination method of this application, wherein:

[0036] The feedback matrix and the resident demand feedback vector are constructed by monitoring network signaling and active reporting of the resident terminal device and collecting feedback data;

[0037] After the record information is forwarded, actual propagation path, depth and node coverage range in the target community network are quantified as an actual coverage utility matrix, and the dimension of the actual coverage utility matrix corresponds to the expected coverage utility matrix;

[0038] The explicit interaction signals of the resident terminal to the information are collected, including clicking, reading time, scoring, forwarding or subscribing operation, and are aggregated and normalized into a resident demand feedback vector, and the resident demand feedback vector is semantically associated with the row dimension of the community feature matrix.

[0039] As a preferred scheme of the network information propagation method of the application, wherein:

[0040] The difference between the actual coverage utility matrix and the expected coverage utility matrix is calculated, and the feedback matrix is combined to update the forwarding utility model and the community feature matrix, including:

[0041] The difference between the actual coverage utility matrix and the expected coverage utility matrix is calculated, and the feedback matrix is combined to update the forwarding utility model and the community feature matrix, including:

[0042] The prediction error matrix is used as training data together with the resident demand feedback vector, and the parameters of the forwarding utility model are updated through online learning;

[0043] The resident interest change reflected by the resident demand feedback vector is fused into the community feature matrix of the target community in incremental learning to update the estimation of the information feature probability distribution.

[0044] The application provides a network information propagation system, comprising:

[0045] A community feature management module is configured to set a community feature matrix for each community, extract a feature vector of the to-be-propagated information when receiving the to-be-propagated information, calculate the similarity between the feature vector and the column vector of the community feature matrix of the target community, and fuse the feature vector into the current time column of the community feature matrix to obtain an updated community feature matrix, and calculate the entropy value change of the community feature matrix before and after the update as a propagation value weight of the to-be-propagated information for the target community;

[0046] A forwarding decision module is connected with the community feature management module and configured to obtain a community space-time context feature at the current time, and input the propagation value weight and the community space-time context feature into a forwarding utility model to predict the expected coverage utility generated by forwarding the information to each neighbor node in the target community.

[0047] The feedback optimization module is connected with the community feature management module and the forwarding decision module respectively, is used for collecting a feedback matrix constructed by resident feedback, a resident demand feedback vector and an actual coverage utility matrix, performing difference calculation on the actual coverage utility matrix and the expected coverage utility matrix, and updating the forwarding utility model and the community feature matrix in combination with the feedback matrix and the demand feedback vector.

[0048] The beneficial effects of the present application are as follows:

[0049] The present application changes the rough evaluation method of the traditional method which depends on keyword matching or fixed rules by setting the community feature matrix to structure and time sequence the operation of the community information state, and accurately measuring the uncertainty reduction brought by new information to the community by calculating the information entropy change before and after the new information is integrated into the community feature matrix.

[0050] The present application changes the rough evaluation method of the traditional method which depends on keyword matching or fixed rules by setting the community feature matrix to structure and time sequence the operation of the community information state, and accurately measuring the uncertainty reduction brought by new information to the community by calculating the information entropy change before and after the new information is integrated into the community feature matrix.

[0051] The present application also sets an optimization process driven by resident feedback, collects feedback matrix and resident demand feedback vector by establishing feedback matrix and resident demand feedback vector, and performs difference analysis on the output of the forwarding utility prediction model, and simultaneously updates the forwarding utility prediction model and the community feature matrix through online learning, so that the system has continuous self-learning and self-adaptive ability, can track the drift of community interest, the change of resident behavior mode and the network environment in real time, and thus maintains the optimal propagation performance. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Among them:

[0053] Figure 1 The network information propagation method provided by the present application is a network information propagation method overall flowchart;

[0054] Figure 2 The community feature matrix update and propagation value evaluation flowchart of the network information propagation system provided by the present application;

[0055] Figure 3 A forwarding decision and closed-loop optimization flowchart of a network information propagation method provided by the present application;

[0056] Figure 4 A preferred representation example of an expected coverage utility matrix of a network information propagation method provided by the present application;

[0057] Figure 5 A preferred representation example of a prediction error matrix of a network information propagation method provided by the present application;

[0058] Figure 6 A campus information propagation platform of a network information propagation method provided by the present application. DETAILED DESCRIPTION

[0059] In order to make the above objectives, features and advantages of the present application more apparent and comprehensible, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0060] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from the description, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, so the present application is not limited by the specific embodiments disclosed below.

[0061] Secondly, the "embodiment" or "embodiment" referred to herein refers to a specific feature, structure or characteristic that can be included in at least one implementation of the present application. "In an embodiment" appearing in different places in this specification does not all refer to the same embodiment, nor is it an embodiment that is separate or selectively excluded from other embodiments.

[0062] Embodiment 1

[0063] As shown in Figure 1 A network information propagation method, comprising

[0064] S1, the information relay node sets a community feature matrix for each community, when receiving the to-be-propagated information, extracts a feature vector of the to-be-propagated information, calculates the similarity of the feature vector and the column vector of the community feature matrix of the target community, integrates the feature vector into the current time column of the community feature matrix, and obtains an updated community feature matrix;

[0065] The behavior information feature dimension of the community feature matrix, the column is the time dimension, and the matrix element represents the probability of the feature appearing in the corresponding time window;

[0066] The community feature matrix set for each community includes defining the row of the community feature matrix as an information feature dimension, which is used to represent different types of information content attributes;

[0067] The columns of the community feature matrix are defined as time windows, used to represent the time series evolution of the community information state.

[0068] Each element value in the community feature matrix is the probability estimate of its corresponding feature appearing or spreading in the community within the corresponding time window.

[0069] The community feature matrix is updated in a fixed-size sliding time window manner over time: the window size is preset as T, when a new time window arrives, the earliest 1 time window column is deleted, and the current time window column is added, ensuring that the number of matrix columns remains fixed.

[0070] For each community covered by the information relay node, the information relay node creates and maintains a data structure to represent the evolution of the community information state, which is the community feature matrix.

[0071] In this application, a preferred implementation method of the community feature matrix includes:

[0072] The community feature matrix is a two-dimensional table structure, and the rows are defined as information feature dimensions. Each row corresponds to a pre-defined information content attribute or feature type, and the feature type constitutes a feature set used to describe the core content of the information, for example, the feature can be the topic category of the information, such as community announcements, neighborhood mutual assistance, and commodity promotion, etc. Key word labels extracted from the information content.

[0073] The columns of the community feature matrix are defined as time dimensions, and each column corresponds to a discretized, continuous time window, such as a time period in hours, days, or weeks. The columns are arranged in chronological order to form a time series, which is used to reflect the historical trajectory of the community information state.

[0074] A preferred meaning and method of obtaining the elements of the community feature matrix includes:

[0075] The element in the community feature matrix located in the ith row corresponding to the ith information feature and the jth column corresponding to the jth time window, the numerical value represents a probability estimate value, which is used to represent the frequency of information with the ith feature appearing or spreading in the current community within the jth time window.

[0076] Specifically, the probability estimate can be obtained by statistical methods, for example, the information relay node can count the total number N of information belonging to the target community that flows through or is observed by the node within the current time window, and then calculate the number of information ni whose content matches the ith feature, matrix element M[i][j]=ni / N, ensuring that the sum of column vector elements is 1.

[0077] In the present application, the information relay node establishes a structured, time-varying information portrait for each community, which records the past reactions of the community to different content information.

[0078] By keyword extraction, topic classification and embedding representation on the to-be-propagated information, a normalized feature vector aligned with the community feature matrix row dimension is generated;

[0079] After completing the establishment and maintenance of the community feature matrix, the information relay node has a benchmark to evaluate the value of new information. When the information relay node receives a to-be-propagated information, the matching degree of the to-be-propagated information with the current information state of the target community is quantified.

[0080] Specifically, the feature extraction and vectorization representation of the to-be-propagated information include keyword extraction, topic classification and embedding representation

[0081] The keyword extraction identifies the most representative and discriminative core words or phrases from the text information as features through natural language processing.

[0082] In the present application, a preferred implementation method of keyword extraction includes:

[0083] The information relay node first pre-processes the text content in the received to-be-propagated information, including: performing word segmentation processing on continuous text to divide it into independent lexical units; removing stop words with extremely low contribution to information theme representation, such as common virtual words and high-frequency general-purpose words; selectively performing stem extraction or morphological reduction on words to merge different forms of the same word into a basic form.

[0084] After pre-processing, the information relay node extracts keywords from the processed word set using a statistical feature-based method.

[0085] A preferred implementation example is the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm.

[0086] Specifically: the term frequency TF is calculated by counting the frequency of each candidate word appearing in the current to-be-propagated information text.

[0087] The inverse document frequency IDF of each candidate word is calculated through a pre-constructed or dynamically maintained background corpus. The inverse document frequency is used to measure the general importance of the word in the entire corpus. The more widely a word appears, the lower the inverse document frequency value.

[0088] The term frequency TF value and the inverse document frequency IDF value of each candidate word are multiplied to obtain its TF-IDF comprehensive score, which is used to quantify the prominence of the word in the current text and its specificity in the entire language background.

[0089] Further, keyword screening is performed, and all candidate words are ranked according to the TF-IDF comprehensive score, and the top N words or words with scores exceeding a certain preset threshold are selected as the final extracted keywords.

[0090] The topic classification determines the pre-defined topic category to which the information belongs through a classification model, such as government affairs, literary style, “life service, etc. Each category can be regarded as a feature.

[0091] The embedding representation adopts a word vector model to map the whole or part of the content of the information into a low-dimensional, dense real number vector, and each dimension of the real number vector carries semantic information.

[0092] Further, the extracted original feature representation, such as the keyword list, the topic label, the original embedding vector, etc. is integrated and normalized to form a fixed-length feature vector, and each dimension of the feature vector corresponds to a row of the community feature matrix, that is, a pre-defined information feature dimension, and the value is used to represent the intensity of the information to be propagated in the feature dimension.

[0093] From the community feature matrix, the column vector of the current time window is selected as the target vector;

[0094] Specifically, a preferred method for selecting the target community current state vector includes:

[0095] In order to evaluate the matching degree of the information to be propagated and the current state of the community, the information relay node selects the column vector representing the latest time window from the community feature matrix corresponding to the target community, which completely records the propagation probability distribution of the community to various information features in the latest statistical period. The propagation probability distribution is regarded as a quantitative snapshot of the current information interest of the community, which is used as a benchmark for comparison and is called a target vector.

[0096] The cosine similarity between the information feature vector and the target column vector is calculated as a similarity measure.

[0097] Specifically, after obtaining the normalized feature vector representing the new information and the target vector representing the current state of the community, the information relay node calculates the similarity between the information feature vector and the target column vector. A preferred cosine similarity is used as a measurement standard. The cosine similarity measures consistency by calculating the cosine value of the angle between the information feature vector and the target column vector in the direction. The range of the cosine value is within a preset range. The closer the value is to the upper limit of the preset range, the more consistent the direction of the information feature vector and the target column vector, that is, the more similar the information feature distribution and the current interest distribution of the community. The closer the value is to the lower limit of the preset range, the greater the difference between the information feature vector and the target column vector. The similarity measurement is used to reflect the degree of consistency between the content of the information to be propagated and the recent information environment of the target community.

[0098] The specific implementation method of the preset range includes:

[0099] In calculating the cosine similarity, the preset range is determined by the inherent properties and data characteristics of the normalized feature vector and the target column vector. The theoretical value range or the actual effective value range can be determined in advance.

[0100] Specifically, the mathematical definition of the cosine similarity determines that its theoretical output value range is [-1, 1]. Among them:

[0101] 1 indicates that the directions of the two vectors are completely the same.

[0102] -1 indicates that the directions of the two vectors are completely opposite.

[0103] 0 indicates that the two vectors are orthogonal, that is, unrelated.

[0104] In the specific application scenario of the present application, the preset range is preset according to the non-negativity constraint of the vector elements and the actual data distribution:

[0105] Specifically, the non-negativity constraint of the vector elements means that all element values are non-negative. When all components of the two vectors are non-negative, the directional cosine value of the angle between them cannot be negative. The actual reachable value range of the cosine similarity is shrunk from the theoretical [-1, 1] to [0, 1].

[0106] The actual data distribution, if a specific encoding or normalization method is used for the feature vector, for example, using zero-mean encoding, theoretically, negative elements can be generated. At this time, according to the specific feature engineering method used, the effective range of the actual data distribution is determined in advance through historical data statistics, for example, [-0.5, 1] or [0.2, 0.95], etc.

[0107] When the preset range is [0, 1], the similarity result can be directly explained as a continuous measure from no correlation to complete consistency, and the similarity value is converted into 0 to 1 through a monotonic non-decreasing mapping relationship

[0108] The present application converts the unstructured information content to be propagated into a structured and computable vector, and realizes objective quantitative evaluation of the matching degree of information and community environment through similarity comparison with a continuously updated community state matrix.

[0109] The similarity calculation fusion weight is used to perform interpolation calculation on the information feature vector and the current time column vector of the community feature matrix, to obtain an updated current time column vector.

[0110] A preferred specific implementation method of the similarity calculation fusion weight includes:

[0111] The similarity measure, i.e., the cosine similarity value, is converted into a fusion weight.

[0112] The fusion weight is a value between 0 and 1, and the specific value is determined by the similarity value through a preset monotonic non-decreasing mapping relationship. The higher the similarity, the more consistent the new information is with the current interest of the community, and the greater the corresponding fusion weight should be in updating the community state; on the contrary, the lower the similarity, the smaller the proportion should be, ensuring that the updating strength of the community feature matrix is proportional to the relevance of the new information and the community environment.

[0113] Further, the information relay node performs calculation of the update operation, taking the fusion weight as a control parameter, to perform interpolation operation on the information feature vector and the target column vector.

[0114] Specifically, the normalized feature vector representing the new information is combined with the current time column vector, i.e., the target vector, extracted from the community feature matrix.

[0115] A preferred example: using linear interpolation method, the updated target vector value is equal to the result obtained by adding the new information feature vector in the proportion of the fusion weight and the current time column vector in the proportion of 1 minus the fusion weight.

[0116] Through linear interpolation, the features of the new information are labeled to the description of the current state of the community with controllable strength: if the information is highly relevant, the community state vector will be significantly close to it; if the relevance is weak, the community state vector will only be fine-tuned. The evolution of the matrix can absorb new knowledge and maintain the continuity of the historical state, avoiding sudden changes due to a single piece of information.

[0117] Specifically, after the similarity is calculated and the fusion weight is obtained, the information relay node performs an update operation on the target community feature matrix, for integrating the influence of the new information into the representation of the community state in a controlled manner.

[0118] Input: Vnew is the normalized feature vector of the information to be propagated, i.e. length N, N equals the number of rows of the community feature matrix, i.e. the number of feature dimensions.

[0119] Vcurrent is the current time column vector extracted from the target community feature matrix, also with length N.

[0120] α is the fusion weight based on the similarity calculation, a scalar value between 0 and 1.

[0121] Output: Vupdated is the updated current time column vector, also with length N, used to replace the original column in the matrix.

[0122] Further, the update is realized through a weighted average algorithm, and its calculation formula is essentially a linear interpolation. The specific calculation is performed element by element according to the feature dimensions. For the i-th feature dimension, i from 1 to N, the updated vector value Vupdated[i] is determined by the following formula: Vupdated[i] = (1-α)*Vcurrent[i] + α*V_new[i]

[0123] Between preserving the historical state Vcurrent and introducing new information Vnew, a balance point is sought according to the matching degree α of the information.

[0124] When the information is highly relevant and α tends to 1, the (1-α) term in the calculation formula approaches 0, while the α term approaches 1, so the contribution of Vcurrent is greatly weakened, and the contribution of Vnew dominates. The updated community state vector Vupdated will approach the feature distribution of the new information, quickly absorbing this highly relevant new knowledge.

[0125] When the information relevance is weak and α tends to 0, the (1-α) term in the calculation formula approaches 1, while the α term approaches 0, so V_current is preserved as it is, and the influence of Vnew is negligible. The community state vector Vupdated only undergoes extremely small adjustment relative to Vcurrent, maintaining the stability and continuity of the community interest state.

[0126] After the calculation of all N dimensions is completed, the complete updated vector Vupdated is generated, and the information relay node performs a replacement operation: Vupdated covers the data originally representing the current time window in the community feature matrix. At this point, the community feature matrix completes a controlled real-time update for a single piece of information.

[0127] Through this specific and quantitative updating mechanism, the system realizes the timely reflection of the influence of new information in the community model; through the adjustment of the fusion weight a, the updating strength is ensured to be proportional to the information relevance, and unreasonable and violent disturbance of the long-term portrait of the community caused by a single accidental or low-correlation information is avoided.

[0128] The updated column vector is replaced in the corresponding column in the original community feature matrix to update the community feature matrix.

[0129] After calculating the similarity between the feature vector of the to-be-propagated information and the current state vector of the target community, the information relay node updates the community feature matrix based on the similarity of the new feature information contained in the to-be-propagated information.

[0130] In the present application, a preferred specific implementation method of iterative replacement of the community feature matrix includes:

[0131] After the interpolation calculation is completed, the updated current time column vector is obtained, and the information relay node directly replaces the column originally representing the current time window in the original community feature matrix through the updated current time column vector, and the community feature matrix of the target community is updated once for the to-be-propagated information.

[0132] In the present application, the column vector of the latest time window in the community feature matrix no longer only contains original statistical data in the past time window, but also fuses new information features that are newly received and evaluated to be valuable, so that a more accurate and timely representation of the real-time information state of the community is realized.

[0133] The iterative replacement of the community feature matrix makes the static data statistical table change into a dynamic and learning community information model, and through continuous evaluation and fusion iteration, the community feature matrix can track the migration and evolution of community interest hotspots.

[0134] S2 calculates the entropy value of the updated community feature matrix, and takes the change of the entropy value, i.e., the information gain, as the propagation value weight of the to-be-propagated information for the community; as Figure 2 As shown, the feature vector is extracted from the received to-be-propagated information; the current state vector is obtained from the community feature matrix; the similarity of the two vectors is calculated and converted into a fusion weight a; the matrix column is updated using a weighted average formula; the entropy values of the matrix before and after the update are calculated to obtain a weighted total entropy; the information gain ΔH is calculated, and the propagation value weight is generated.

[0135] Each column of the community feature matrix is regarded as a discrete probability distribution vector describing the distribution of community information features in the time window;

[0136] Specifically, each element value of each column of the community feature matrix represents the probability estimation of the corresponding feature appearing in the corresponding time window. To ensure that it constitutes a legal probability distribution vector, it needs to be verified and normalized before calculation. Specifically, read all element values of the target column, check whether each element satisfies: first, all values are non-negative; second, the sum of all values is 1.

[0137] If the sum is not strictly 1 due to statistical error or calculation accuracy, normalization operation needs to be performed: divide each element value of the current column by the sum of all element values of the column to obtain the standardized probability distribution vector.

[0138] For the community feature matrix before and after updating respectively, calculate the Shannon entropy of each probability distribution vector by column to obtain the entropy value sequence arranged by time window;

[0139] Specifically, calculate the Shannon entropy of each column for each probability distribution vector, i.e. each column of the matrix. The calculation method of the Shannon entropy is: for each probability value greater than zero in the given probability distribution, calculate the inverse of the product of the natural logarithm of the probability value and the probability value itself, and then accumulate and sum all such inverse products in the current column. The result obtained is the Shannon entropy value of the distribution. This calculation process is applied to each column of the two community feature matrices before and after updating respectively.

[0140] After obtaining the Shannon entropy, two numerical lists are generated, and the number of values in each list is equal to the number of columns of the matrix, i.e. the number of time windows.

[0141] A preferred method for calculating Shannon entropy includes: .

[0142] wherein, H(j) is the Shannon entropy of the jth time window; M[i][j] is the probability value of the ith row and jth column of the community feature matrix; : the minimum value avoids meaningless ln(0); H(j) ≥ 0, the larger the entropy value, the more dispersed the information feature distribution, and the higher the uncertainty.

[0143] Single time window Shannon entropy is used to quantify the uncertainty of the community information state in a single time window.

[0144] The values in each list are arranged according to the column sequence of the corresponding time window in the matrix.

[0145] Specifically, the entropy value corresponding to the earliest time window is arranged at the starting position of the sequence, the entropy value of the second earliest time window follows it, and so on, and the entropy value of the latest time window is arranged at the end of the sequence. The two numerical lists arranged in time sequence, i.e. the entropy value sequence arranged by time window.

[0146] The entropy values of each time window are summed up by time weight, and the weighted total entropy of the community feature matrix before and after updating is obtained respectively.

[0147] Specifically, the system defines a time weight sequence in advance, and the time weight sequence has the same length as the entropy value sequence, that is, the same number of time windows, and each weight value corresponds to a time window.

[0148] A preferred example of the basic principle of weight setting is: the more recent the window, the higher the importance, and the larger the weight value allocated; the more distant the window, the smaller the weight value. A preferred implementation example is to use exponential decay weighting, that is, the weight value decreases exponentially with time backtracking.

[0149] When calculating the weighted total entropy, each entropy value in the entropy value sequence is multiplied by the weight value in the same order position in the time weight sequence to obtain a series of weighted entropy values.

[0150] Further, all weighted entropy values are added to obtain the sum of entropy, which is the weighted total entropy of the community feature matrix before or after updating. The entropy value sequence corresponding to the matrix before and after updating is executed once respectively, and two weighted total entropy values are obtained.

[0151] The entropy value change is the difference between the weighted total entropy before updating and the weighted total entropy after updating. The difference value is used as the information gain brought by the target community to the information to be propagated. The greater the entropy reduction, the higher the value, and directly or after normalization processing, as the propagation value weight of the information to be propagated to the community.

[0152] After obtaining two weighted total entropies, the weighted total entropy after updating is subtracted from the weighted total entropy before updating to obtain the information gain, which is used to quantify the degree of change in the overall information state distribution of the community caused by the integration of new information.

[0153] The information gain difference can be used as a propagation value weight after normalization processing. A preferred implementation is to set the maximum possible information gain reference value based on historical data or theoretical analysis to a positive value, take the actual information gain difference value as non-negative, divide by the maximum possible information gain reference value, and map the weight to between 0 and 1. It should be noted that the entropy reduction is a positive weight, and the entropy increase is a 0 weight.

[0154] The community feature matrix row corresponds to semantic features, the column corresponds to the sliding time window, and the element value is the probability of the feature appearing in the window. The community information state is promoted from a discrete event sequence to a continuous probability evolution model, realizing the quantization, timing, and structured memory of community interest. The cosine similarity calculation fuses the weight a, and the current time column is updated by weighted interpolation, realizing the adaptive smooth fusion of new information and community state, avoiding information dilution of statistical average and preventing mutation caused by single abnormal information, ensuring the robust evolution of community portrait. Each column of the matrix is regarded as a probability distribution vector, the weighted total entropy difference before and after updating is calculated, the information gain is directly converted into propagation value weight, the entropy change principle of information theory is embedded in matrix operation, forming an objective and continuous value evaluation model, replacing subjective rules or simple keyword matching.

[0155] S3, the information relay node obtains the community spatio-temporal context feature at the current time, inputs the propagation value weight and the community spatio-temporal context feature into a forwarding utility model, and predicts the coverage utility of the information forwarded by each neighbor node in the target community;

[0156] Inputting the propagation value weight and the community spatio-temporal context feature into the forwarding utility model comprises:

[0157] The information relay node obtains the real-time geographic position and timestamp of the node itself through the built-in positioning and timing module, obtains the community belonging identifier, real-time position, moving speed and direction through the exchange of signaling with neighbor nodes.

[0158] The feature vector composed of the propagation value weight and the obtained context feature is input into the forwarding utility model.

[0159] Specifically, the information relay node constructs a spatio-temporal context feature set for describing the current community environment through local perception and neighbor cooperation. The spatio-temporal context feature set includes: node self-state, neighbor node dynamic information, and community activity and historical pattern.

[0160] In the node self-state, the information relay node obtains the real-time geographic position coordinates of itself through the built-in positioning module, and obtains the timestamp through the timing module. The timestamp not only contains absolute time, but also can be associated with social time attributes such as weekdays and holidays.

[0161] A preferred positioning module can be a global satellite navigation system GNSS receiver or an indoor positioning unit of a wireless signal

[0162] The neighbor node dynamic information is used to understand the local network topology and node flow mode. The specific information is periodically exchanged by the information relay node and the neighbor node in the communication range through the exchanged signaling data packet. The node can obtain the community belonging identifier and real-time position and movement state and connection quality index.

[0163] Specifically, in the community belonging identifier, the community ID mainly associated or served by the neighbor node is used to determine whether it belongs to the target community or the related community.

[0164] In the real-time position and movement state, the current position coordinates, instantaneous movement speed and movement direction of the neighbor node. These information can be used to infer the movement mode of the node and the trend of the relative position relationship with the information relay node or the geographical area of the target community.

[0165] A preferred connection quality index also includes wireless link quality information of the neighbor node, such as signal strength, signal-to-noise ratio, etc., which is used to evaluate the reliability of the forwarding link.

[0166] In the community activity and historical mode, the information relay node can query the static or semi-static knowledge stored locally or obtained from the community management platform, such as community activity schedule and historical flow mode, for example, the flow density of the community entrance is high in the evening of weekdays. By matching the current time and geographical position with the knowledge base, combined with time weight fusion, for example, the real-time data weight is 0.7 and the historical mode weight is 0.3, the current activity state or typical flow stage of the community is inferred, wherein the weight of the historical mode is exponentially attenuated with the distance from the current time.

[0167] After obtaining the multi-source heterogeneous data of the community belonging identifier and the real-time position and movement state and the connection quality index, the node performs data fusion and structured processing, and converts it into an input format suitable for the forwarding utility model processing.

[0168] Further, the forwarding utility model establishment method includes constructing a feature vector, model input, predicting actual propagation effect, and forwarding utility model optimization and model training.

[0169] In the construction of the feature vector, the information relay node creates a structured fusion feature vector. The fusion feature vector is the propagation value weight, and all collected spatio-temporal context features are standardized and coded as input data of the forwarding utility model.

[0170] Specifically, the geographical position can be converted into a coordinate offset relative to a community center point.

[0171] The timestamp can be coded as a classification feature such as time of day, day of week, etc.

[0172] The information of the neighbor nodes is counted, and a feature sub-vector is formed for each important neighbor, and then the whole is pooled.

[0173] The community activity state can be represented as a Boolean value.

[0174] The model input includes a fusion feature vector of information value and multi-dimensional description of the community space-time environment, which is used as input data and fed into the forwarding utility model deployed in advance on the node.

[0175] The forwarding utility model is used to evaluate the expected revenue that can be brought by taking different forwarding actions on different value information under different environment states.

[0176] The forwarding utility model takes the fusion feature vector as input, predicts the expected information coverage increment and propagation efficiency gain in the target community after forwarding information to different neighbor nodes of the information relay point within a future preset time window, and takes the expected information coverage increment and the propagation efficiency gain as output targets.

[0177] The initial training of the forwarding utility model is based on a historical propagation record data set, and the data set includes feature input and actual propagation effect sample pairs.

[0178] It should be noted that the initial training of the forwarding utility model is based on a historical propagation record data set, and the data set is divided according to a time window, and each sample pair includes a fusion feature input corresponding to the time window and an actual propagation effect within a preset time after the window, so as to ensure that the time dimension of the training data is aligned with the time window of the model prediction.

[0179] The forwarding utility model is established by learning the nonlinear relationship in the historical data.

[0180] In this application, a preferred construction method of the forwarding utility model includes:

[0181] The forwarding utility model is a prediction function for evaluating the future revenue that may be brought by a forwarding action, and the forwarding utility model takes the fusion feature vector of the current time as input, and performs independent utility prediction for each candidate neighbor node of the information relay node.

[0182] Further, the forwarding utility model outputs one or a group of prediction values for each candidate neighbor node, which are used to quantify the effect that can be achieved in the target community within a future preset time window if the current to-be-propagated information is selected to be forwarded to the neighbor node. The prediction output targets include:

[0183] The expected information coverage increment predicts how many percentages of the target community residents or nodes will first receive the information or how much the geographical topological coverage range of the information in the community network will be expanded by the forwarding action.

[0184] The propagation efficiency gain is expected, and the prediction forwarding action will shorten the average time required for information to propagate to key nodes in the community or reach a certain coverage rate, or reduce the number of redundant message copies generated during the entire propagation process, and improve resource utilization efficiency.

[0185] The training data set of the forwarding utility model includes feature input and label actual effect.

[0186] The training data set is constructed by collecting a historical propagation record data set, which is constructed by long-term recording of actual propagation events in the network, and each record constitutes a sample pair, wherein:

[0187] The feature input corresponds to a certain time in history when a node receives a piece of information, and the constructed fusion feature vector contains the information value weight and the spatio-temporal context feature at that time.

[0188] The label actual effect part records the actual forwarding action taken by the node in the current historical scenario, such as selecting a certain neighbor for forwarding, and the change in coverage rate and propagation efficiency index of the information in the target community within a certain period of time thereafter.

[0189] The forwarding utility model is trained by using the training data set, and the forwarding utility model excavates and establishes a complex and nonlinear mapping relationship between the fusion feature vector, the forwarding action and the future propagation effect from the data.

[0190] A preferred fusion feature vector calculation method is:

[0191] ;

[0192] Wherein, The community spatio-temporal context feature includes node position, timestamp, neighbor moving state, etc. () is a feature encoding function, i.e. heterogeneous context features are converted into vectors; Concat() is a vector concatenation operation;

[0193] ∈ The fusion feature vector is d, and the model input dimension is d. The propagation value weight ∈ [0, 1], the greater the propagation value weight value, the higher the information priority.

[0194] ;

[0195] Wherein, ΔH is the difference between the weighted total entropy before and after updating, i.e. information gain, wherein, Max() is the maximum information gain, Norm() is a normalization operation, and max is the maximum value.

[0196] A preferred implementation method of the forwarding utility model includes:

[0197] The input layer serves as the model entrance, and the number of neurons corresponds to the dimension of the fused feature vector. The input layer is used to receive and normalize the input fused feature vector.

[0198] The feature abstraction and interaction layer, i.e., one or more hidden layers, is connected to the input layer. Each of the multiple fully connected hidden layers is composed of a number of neurons, which are used to learn and extract abstract features and complex feature interaction relationships from the original fused features through linear transformation and nonlinear activation functions. For example, the first layer learns the combination pattern of geographical location and time features, and deeper layers learn the deep associations between information value weights, neighbor node movement patterns, and community activity states.

[0199] The activation function enables the forwarding utility model to capture and model complex nonlinear relationships between inputs and outputs.

[0200] Neighbor node attention layer: To further improve the accuracy of the decision, the forwarding utility model can introduce an attention mechanism layer, i.e., a neighbor node attention layer, after the hidden layer.

[0201] Specifically, the forwarding utility model does not treat all potential neighbor node-related features equally. Instead, it dynamically calculates attention weights, which represent the degree of attention paid by the forwarding utility model to different neighbor nodes or the importance of different environmental feature dimensions to the final decision under the current specific fused feature input.

[0202] The last layer of the forwarding utility model is the output layer, which has the same number of neurons as the number of candidate neighbor nodes that the information relay node needs to evaluate at the moment. Each candidate neighbor corresponds to one or two output neurons. The output layer is used to map the features represented after depth processing and abstraction in the previous layers to the final quantized prediction value for each neighbor node. If a linear activation function is used, the output value represents the predicted absolute utility score. If a function such as Sigmoid is used, the score can be normalized to a fixed interval.

[0203] The prediction process and output data of the forwarding utility model specifically include:

[0204] After the forwarding utility model is built and trained, its prediction process is a forward propagation process: the information relay node inputs the currently constructed fused feature vector into the model, and the data flows through the input layer, the feature abstraction and interaction layer, and the attention layer for focused processing, and finally reaches the output layer.

[0205] After the output layer activation, a prediction utility vector is generated, each element in the prediction utility vector corresponds to a predicted forwarding utility value of a specific candidate neighbor node. The predicted forwarding utility value is a prediction scalar, which encodes the estimation of the expected information coverage increment and propagation efficiency gain that the forwarding utility model can bring to the neighbor after forwarding information to the neighbor within a preset future time window. The information relay node can identify which neighbors to forward to maximize the propagation benefit of the information by comparing all utility values in the prediction utility vector under the current scenario.

[0206] S4, collect the feedback matrix and the resident demand feedback vector constructed by the resident feedback, difference calculate the feedback matrix and the expected coverage utility matrix, and update the forwarding utility model and the community feature matrix.

[0207] As shown in Figure 4 , in an example of the expected coverage utility matrix, is the forwarding utility model; θ is the model parameter trained based on historical data; n is the number of candidate neighbor nodes; is the expected coverage increment to the first neighbor; is the expected propagation efficiency gain to the first neighbor; ∈ is the expected coverage utility matrix, and n is a positive integer.

[0208] By monitoring network signaling and active reporting of resident terminal devices, feedback data is collected to construct a feedback matrix and a resident demand feedback vector;

[0209] After the record information is forwarded, the actual propagation path, depth and node coverage range in the target community network are quantified into an actual coverage utility matrix, and the dimension of the actual coverage utility matrix corresponds to the expected coverage utility matrix;

[0210] The explicit interaction signal of the resident terminal to the information can include clicking, reading time, rating, forwarding or subscribing operation, and is normalized into a resident demand feedback vector. The row dimension of the community feature matrix is semantically associated with the resident demand feedback vector, and each weight can be set according to the actual community resident group, for example, the click weight is 0.3, the reading time weight is 0.2, the rating weight is 0.3, the forwarding weight is 0.15, and the subscription weight is 0.05. After weighted summation according to the preset weight, the resident demand feedback vector is normalized to the value range [0, 1].

[0211] It should be noted that the resident demand feedback vector and the row dimension of the community feature matrix are semantically corresponding: the i-th dimension of the demand feedback vector corresponds to the information feature of the i-th row of the community feature matrix, such as a keyword or a topic classification, and the value of the vector element is the interest intensity of the residents to the information feature, ensuring the semantic consistency of the dimensions.

[0212] The information relay node collects two types of key feedback data from the network and user terminals by combining active monitoring with passive reception:

[0213] The information relay node records the actual diffusion of the information after being forwarded in the target community network by monitoring network layer signaling and tracking the propagation path of information identifiers such as message IDs. The specific data recorded include but are not limited to: the information has reached the information relay node, i.e. the node coverage range, the information has propagated through multiple hops, i.e. the propagation depth, and the main propagation path topology formed.

[0214] Further, the information relay node quantizes and integrates the discrete propagation observation data into an actual coverage utility matrix.

[0215] The actual coverage utility matrix is a two-dimensional data structure, and the rows of the actual coverage utility matrix represent spatial dimensions, specifically corresponding to geographical grid cells formed after dividing the target community or corresponding to node groups defined according to network topology or management logic.

[0216] The columns of the actual coverage utility matrix represent the time dimension, corresponding to the observation period after forwarding the information divided into consecutive feedback statistical time sub-windows.

[0217] Each element of the actual coverage utility matrix is uniquely determined by the i-th spatial region or group corresponding to the row i and the j-th time sub-window corresponding to the column j, and the element stores the actual coverage utility measurement value quantized. The value is an indicator of the actual propagation effect of the target information in the i-th spatial region or node group in the j-th time sub-window, obtained by monitoring and statistics.

[0218] The information relay node also receives active reporting or interaction logs from resident terminal devices. That is, explicit interaction signals, such as: information click opening behavior, reading page dwell time, information scoring, secondary forwarding behavior, or information subscription operations, etc.

[0219] After the information relay node collects the explicit interaction signals, it performs cleaning, aggregation and normalization processing. For example, all resident scores for the information are averaged, or the click rate and reading time are weighted and combined, and finally the processed explicit interaction signals are integrated to form a resident demand feedback vector.

[0220] Each dimension of the resident demand feedback vector is semantically associated with a row dimension of the community feature matrix, i.e., an information feature dimension, and the value reflects the intensity of interest or satisfaction of the resident group for each feature dimension carried by the information. For example, an information about a community badminton match may generate higher feedback values in the sports and community activity feature dimensions.

[0221] After obtaining the resident demand feedback vector, an optimization process is started, including difference calculation and error analysis and online updating of the forwarding utility model.

[0222] The difference calculation and error analysis compares the actual coverage utility matrix with the expected coverage utility matrix element by element or based on overall indicators, for example, calculates the difference between the corresponding position elements of the two matrices, generates a plurality of quantified prediction error indicators by comparison, and forms a prediction error matrix.

[0223] As shown in the following figure, starting from the propagation value weight; collecting various spatio-temporal context features; constructing a fusion feature vector; predicting an expected coverage utility matrix through a forwarding utility model f_θ; executing a forwarding decision according to the prediction result; collecting implicit and explicit feedback data; calculating a prediction error for updating the model; updating the community feature matrix using resident feedback; and the system enters the next round of preparation state. Figure 3 The online updating of the forwarding utility model uses the prediction error matrix and the original feedback data used to construct the prediction error matrix as a supervision signal, combines the original fusion feature vector input, and uses a small batch update of a stochastic gradient descent to perform an iterative adjustment of the model parameters of the forwarding utility model. The forwarding utility model corrects itself by minimizing the prediction error, so that the future prediction of the forwarding utility model is closer to the real network propagation law.

[0224] The resident demand feedback vector is used to directly update the community feature matrix of the target community.

[0225] Specifically, the resident demand feedback vector is regarded as a sampling survey result of the interest of the residents in the current latest time window, and is fused into the current time column of the community feature matrix, so that the matrix not only based on the frequency of information propagation, but also can integrate signals reflecting the subjective preference and satisfaction of the residents, so as to make the representation of the community information state more comprehensive.

[0226] The difference calculation between the actual coverage utility matrix and the expected coverage utility matrix, and the updating of the forwarding utility model and the community feature matrix combined with the feedback matrix, include:

[0227] Calculate the difference between the actual coverage utility matrix and the expected coverage utility matrix to obtain a model prediction error matrix;

[0228]

[0229] ​The prediction error matrix and the resident demand feedback vector are taken as training data together, and the parameters of the forwarding utility model are updated through online learning;

[0230] The update of the community feature matrix follows the principle of immediate update priority: first, complete the immediate fusion update of the information feature to be propagated through S1, and then in S4, the resident demand feedback vector is integrated into the current time column vector in an incremental learning manner. A preferred example: new column vector value = 0.7*S1 updated column vector + 0.3*resident demand feedback vector, update the estimate of the information feature probability distribution.

[0231] The information relay node first evaluates the prediction accuracy, specifically, the expected coverage utility matrix output by the forwarding utility model is directly compared with the actual coverage utility matrix constructed by actual observation, the comparison method is in matrix unit, the specific method is to calculate the numerical difference, such as difference or relative error, of the elements in the same row and column position of the two matrices, that is, corresponding to the same space area and time sub-window.

[0232] Further, the calculation results of all positions are also filled into a new matrix, which is the model prediction error matrix. The model prediction error matrix is used to represent the overestimation or underestimation of the forwarding utility model in predicting the coverage utility of information in different parts of the community and different time periods.

[0233] Further, after obtaining the error matrix, the forwarding utility model is optimized.

[0234] Specifically, the fusion feature vector as the input of the model, the expected coverage utility matrix generated by the model decision, and the newly calculated model prediction error matrix jointly constitute an online training sample.

[0235] A preferred prediction error matrix, E[i][c]=| |, wherein is the actual propagation effect of the cth dimension after forwarding to the ith neighbor node, E is the prediction error matrix, E[i][c] is the absolute error value of the ith neighbor node in the cth propagation effect dimension, i is the serial number of the candidate neighbor node, c is the evaluation dimension of the propagation effect, is the actual coverage utility matrix, is the expected coverage utility matrix, is the expected propagation effect of the cth dimension after forwarding to the ith neighbor node, as shown in Figure 5 , is the actual coverage rate increment of the ith neighbor node, is the expected coverage rate increment of the ith neighbor node is the actual propagation efficiency gain of the ith neighbor node Expected propagation efficiency gain for the i-th neighbor node.

[0236] Specifically, each row corresponds to one candidate neighbor node, the 1st row = neighbor node 1, the 2nd row = neighbor node 2, …, the n-th row = neighbor node n;

[0237] Each column corresponds to one propagation effect dimension, the 1st column = error of coverage increment, the 2nd column = error of propagation efficiency gain;

[0238] Each cell stores the predicted absolute error of a single neighbor node + a single effect dimension, for example, the element in the 3rd row and the 1st column | ΔC3, actual - ΔC3, pred | represents the deviation between the actual coverage increment and the expected coverage increment after forwarding to the 3rd neighbor node.

[0239] Further, the small batch stochastic gradient descent is used to take the error matrix as the basis for loss calculation, and the internal connection weight parameters of the forwarding utility model are fine-tuned through back propagation to reduce the future prediction error.

[0240] At the same time, the resident demand feedback vector reflecting the subjective preference of the residents is used to refresh the cognition of the forwarding utility model on the interest of the community information, each dimension of the resident demand feedback vector corresponds to the row semantics of the community feature matrix, and the specific implementation method includes: regarding the resident demand feedback vector as a direct sampling of the community interest at the current time; taking a preset, relatively small fusion rate coefficient as a regulation parameter, and adding the resident demand feedback vector to the current latest time column vector of the target community feature matrix in a weighted average manner.

[0241] For example, the new matrix column vector value = (1-fusion rate coefficient) x original column vector value + fusion rate coefficient x corresponding value of the resident demand feedback vector.

[0242] The estimation of the occurrence probability of each information feature in the current time window in the community feature matrix is not only based on the objective frequency statistics of information propagation, but also incorporates the likes and attentions expressed by the residents, so that the record of the community knowledge base on the interest of the residents is more accurate.

[0243] The rows of the expected coverage utility matrix correspond to the candidate neighbor nodes, and the columns correspond to different utility dimensions, so that the forwarding decision is upgraded from single-value comparison to multi-target structured evaluation. By calculating the element-by-element difference between the expected and actual coverage utility matrices, a prediction error matrix is generated, which not only quantifies the deviation in different neighbors and different utility dimensions, but also provides high-dimensional, structured gradient signals for online learning. The construction of the actual coverage utility matrix requires continuous monitoring of the propagation path and node coverage, and the real physical effects of network propagation are injected back into the model optimization, so that the model optimization is more in line with the real situation of each community.

[0244] Embodiment 2

[0245] As Figure 2 shown, a network information propagation system includes:

[0246] a community feature management module including a matrix initialization unit, a feature fusion unit and an entropy calculation unit; the matrix initialization unit is used to create and maintain a community feature matrix for each community, wherein the rows of the matrix correspond to the information feature dimension, the columns correspond to the time dimension, and the matrix elements represent the probability of the corresponding feature appearing in the corresponding time window;

[0247] the feature fusion unit is used to receive the feature vector of the information to be propagated, calculate the similarity between it and the current time column vector in the target community feature matrix, and fuse the feature vector into the current time column by weighted interpolation according to the similarity, to update the community feature matrix;

[0248] the entropy calculation unit is used to regard each column of the community feature matrix before and after updating as a probability distribution and calculate its Shannon entropy, obtain the total entropy of the matrix by weighted summation, and output the change value of the total entropy as the information gain;

[0249] a forwarding decision module including a context awareness unit and a utility prediction unit;

[0250] The context awareness unit is used to obtain and integrate the node's own geographic location, time, neighbor node community affiliation and moving state, and community activity schedule information in real time through the built-in positioning module, neighbor signaling exchange and local knowledge base, to generate a community spatio-temporal context feature vector;

[0251] The utility prediction unit is used to receive the information gain value from the community feature management module and the context feature vector from the context awareness unit, process them through a pre-trained machine learning forwarding utility model, predict the expected coverage utility matrix generated in the target community after the information is forwarded to different neighbor nodes, and output the expected coverage utility matrix to the neighbor nodes;

[0252] a feedback optimization module including a data collection unit and a model update unit;

[0253] The data collection unit is used to monitor the actual propagation path of the network and the explicit interaction behavior of the resident terminal, to construct an actual coverage utility matrix reflecting the actual propagation effect and a resident demand feedback vector reflecting the change of the resident interest;

[0254] The model updating unit is used to calculate the error between the actual coverage utility matrix and the expected coverage utility matrix output by the forwarding decision module, and combines the resident demand feedback vector to simultaneously update the parameters of the forwarding utility prediction model and the community feature matrix in the community feature management module through an online learning algorithm, to realize closed-loop optimization of the system.

[0255] It should be noted that the present application realizes the structured representation and adaptive fusion and update of the community information state through the community feature matrix, the expected coverage utility matrix and the actual coverage utility matrix, and the error matrix, supports the value quantification based on information entropy, realizes the end-to-end update of the user intention to the community portrait through the semantic alignment of the row dimension of the community feature matrix and the resident demand feedback vector, solves the technical problems in the traditional method that the information value is difficult to quantify, the decision ignores the space-time context, and the system lacks adaptive ability, and further provides a hardware basis for the deployment on the edge node with limited resources through the parallelization characteristics of the matrix operation.

[0256] Embodiment 3

[0257] The school community of a certain college is selected, including 3000 students, 180 faculty members, 2800 parents, and a total of 5980 people, and the core information dissemination needs include: campus notice, club activity recruitment, academic lecture, canteen menu update, safety warning, etc.

[0258] The traditional dissemination adopts the flooding forwarding of the school WeChat group and the campus bulletin board mode, and the problems are: the emergency notice and the unimportant information such as the canteen menu are mixed and sent, which is easy to be missed; the activity notice is forwarded during non-working hours, and the opening rate is less than 15%; the proportion of redundant messages is more than 60%, which causes waste of network bandwidth and heavy receiving burden of terminal devices.

[0259] The present application solves the above problems through the community feature matrix and the entropy value quantification, the space-time context fusion and the feedback closed loop, verifies the creativity and technical effect of the scheme.

[0260] The case implementation details include:

[0261] Construction of the school community feature matrix;

[0262] Matrix structure: row=8 information features, emergency notice, regular notice, club activity, academic lecture, canteen menu, safety warning, and school-parent interaction; column=7-day sliding time window, T=1 day, window size 7 days, daily addition of 1 column and deletion of the earliest 1 column.

[0263] Probability calculation: the dissemination frequency of each type of information is counted, for example, the emergency notice appears twice on Monday, and the total information on the day is 5, so the matrix element M[1][1]=2 / 5=0.4.

[0264] Entropy calculation, for example, "rainstorm emergency notice":

[0265] The weighted total entropy of the community feature matrix before updating is Hold=3.2, and the distribution of each type of information is scattered, with high uncertainty.

[0266] After integrating the notification feature vector, the weighted total entropy of the matrix is Hupdated=1.8, the proportion of emergency notifications is increased, and the uncertainty is reduced.

[0267] Information gain ΔH=3.2-1.8=1.4, normalized propagation value weight ωvalue=0.92, highest priority.

[0268] Collecting campus space-time features:

[0269] The time feature is that the high-activity period is from 7:00 to 8:00 (before school) and from 17:00 to 20:00 (after school / parents' work hours).

[0270] The spatial feature is that teachers and staff are concentrated in office buildings, students are concentrated in teaching buildings, and parents are mostly outside the community (requiring mobile terminal push).

[0271] The neighbor node feature is that the class teacher node and the parent committee node are the core forwarding nodes.

[0272] Fusion feature vector Concat(0.92, Encoding(“17:30 period + parent terminal + class teacher node”)), input forwarding utility model.

[0273] Collecting feedback data:

[0274] After the actual coverage utility matrix is forwarded to the 5 core neighbor nodes, i.e. 3 class teachers and 2 parent committees, the coverage rate reaches 98% within 30 minutes, and the propagation time is 8 minutes.

[0275] The parent click rate in the resident demand feedback vector is 92%, the reading time longer than 30 seconds accounts for 85%, and the forwarding rate is 40%, corresponding to the demand value of the emergency notification feature dimension = 0.95.

[0276] Based on the prediction error matrix, the actual coverage rate is 98% compared with the expected 95%, the error is 3%, and the forwarding utility model parameters are adjusted online to improve the prediction accuracy.

[0277] Scene effect verification: emergency notification scene, traditional method: WeChat group screen, parent missed reading rate 35%, only 70% of parents know within 2 hours;

[0278] The application focuses on the 17:30-18:30 parent active period through the class teacher node directional forwarding with a propagation value weight ωvalue=0.95, and the awareness rate is 99% within 30 minutes, and the missed reading rate is 1%.

[0279] Breakthrough indiscriminate flooding, through the value of quantitative and space-time matching, solve the lack of timely delivery of emergency information.

[0280] Club activities recruitment scene, the traditional method for the whole school push, students click rate of 12%, 50 people to recruit new registration;

[0281] The entropy value calculation in this application is ΔH=0.7, the weight ωvalue=0.6, the space-time feature focuses on the middle school period + 16:00-17:00 after-school period and the science and technology teacher node, and is pushed to the middle school parent student terminal, the click rate is 68%, the new registration is 120 people, and the target group accounts for 95%;

[0282] Through the precise correlation of information features, user groups, and space-time scenes, the problem of low precision and resource waste in traditional methods is solved.

[0283] As shown in Figure 6 , the campus intelligent information dissemination platform is used to count all the data in this embodiment. It should be noted that in the embodiment provided in this application, only part of the content in the campus intelligent information dissemination platform is described, i.e., the information related to the technical scheme disclosed in this application, and the data counted is part of the campus groups as experimental data.

[0284] Through 4 weeks of iterative optimization of the community network message push effect of a certain school:

[0285] Week 1: model prediction accuracy 82%, important information coverage 90%;

[0286] Week 4: after 4 feedback closed-loop updates, model prediction accuracy increased to 97%, important information coverage stabilized at 98%, and redundant message rate decreased to 15%;

[0287] Through the application of the technical scheme disclosed in this application in the school, the important information coverage rate is increased by 33%, which proves that the entropy value quantization value solves the problem that the traditional method cannot distinguish the priority of information;

[0288] The time consumption of dissemination is reduced by 68%, which proves that the space-time context and core node forwarding solve the problem that the traditional method has no scene adaptation ability;

[0289] The redundant message rate is reduced by 75%, which proves that the resource on-demand allocation solves the problem that the traditional method has no resource optimization mechanism.

[0290] It is important to note that the construction and arrangement of the application shown in the various exemplary embodiments is illustrative only. Although only two embodiments have been described in detail in this disclosure, persons of ordinary skill in the art having the benefit of this disclosure will readily appreciate that many modifications are possible without making substantial deviations from the novel teachings and advantages of the subject matter described in this application. For example, the size, shape, arrangement, and relative proportions of the various elements, as well as the values of parameters such as temperature, pressure, etc., the mounting arrangements, the use of materials, colors, orientations, etc., can be changed as desired. For example, an element shown as a single configuration can be divided into two or more separate elements, the position of an element can be reversed or otherwise changed, and the nature or number of elements or positions can be modified or changed. Accordingly, all such modifications are intended to be included within the scope of the present application. The order or sequence of any process or method steps can be changed unless otherwise specified or clearly implied by the circumstances. Any reference signs in the description shall not be construed as limiting the scope of the application. Any "device plus function" clauses are intended to cover the structures described herein as performing their claimed function, and also material equivalent thereof. Other substitutions, modifications, changes and omissions can be made in the design, operating conditions and arrangement of the exemplary embodiments without departing from the scope of the present application. Accordingly, the present application is not limited to the particular embodiments described herein, but extends to various modifications that still fall within the scope of the appended claims.

[0291] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of an actual implementation can be described (i.e., those not necessary to enable one to practice the present application, or those not related to the current consideration of the best mode for practicing the present application).

[0292] It is to be understood that the development of the exemplary embodiments of this application can not be limited to the precise constructional arrangements shown in the drawings, and that various modifications can be made to the exemplary embodiments and the arrangement and combination of parts without departing from the scope of the application as recited in the claims.

[0293] It should be noted that the above examples are intended to be illustrative only and not limiting of the present application. Although the present application has been described in detail with reference to the preferred embodiments, those with ordinary skill in the art will appreciate that various modifications can be made without departing from the spirit and scope of the present application, which is defined by the claims.

Claims

1. A method for disseminating network information, characterized in that, include: S1. The information relay node sets up a community feature matrix for each community. When it receives information to be disseminated, it extracts the feature vector of the information to be disseminated, calculates the similarity between the feature vector and the column vector of the community feature matrix of the target community, integrates the feature vector into the current time column of the community feature matrix, and obtains the updated community feature matrix. Each column of the community feature matrix is ​​regarded as a discrete probability distribution vector describing the distribution of community information features within a time window; For the community feature matrices before and after the update, calculate the Shannon entropy of each probability distribution vector column by column to obtain the entropy value sequence arranged by time window; The entropy values ​​of each time window are weighted and summed according to time weights to obtain the weighted total entropy of the community feature matrix before and after the update. The change in entropy value is calculated as the difference between the weighted total entropy before the update and the weighted total entropy after the update. The difference is used as the information gain that the information to be disseminated brings to the target community, and is used directly or after normalization as the dissemination value weight of the information to be disseminated to the community. S2 calculates the entropy value of the updated community feature matrix and uses the change in the entropy value as the weight of the dissemination value of the information to be disseminated to the community. The propagation value weights and community spatiotemporal context features are input into the forwarding utility model, including: The information relay node obtains its own real-time geographical location and timestamp through a built-in positioning and timing module; and obtains community affiliation identifier, real-time location, movement speed and direction by exchanging signaling with neighboring nodes. The feature vector formed by the propagation value weight and the acquired context features is input into the forwarding utility model; S3. The information relay node obtains the community spatiotemporal context features at the current moment, inputs the propagation value weight and the community spatiotemporal context features into the forwarding utility model, and predicts the coverage utility generated by each neighbor node forwarding information in the target community. S4. Collect the feedback matrix and resident demand feedback vector constructed from resident feedback, calculate the difference between the actual coverage utility matrix and the expected coverage utility matrix, and update the forwarding utility model and community feature matrix. By monitoring network signaling and proactive reporting from resident terminal devices, feedback data is collected to construct a feedback matrix and a resident demand feedback vector. After the recorded information is forwarded, the actual propagation path, depth and node coverage in the target community network are quantified into an actual coverage utility matrix, and the dimension of the actual coverage utility matrix corresponds to the expected coverage utility matrix. Explicit interaction signals from resident terminals in response to the information are collected and normalized into resident demand feedback vectors, which are semantically related to the row dimensions of the community feature matrix.

2. The network information dissemination method as described in claim 1, characterized in that: The behavioral information feature dimension of the community feature matrix has a time dimension as the column, and the matrix elements represent the probability of a feature appearing within the corresponding time window. The step of setting a community feature matrix for each community includes defining the rows of the community feature matrix as information feature dimensions to represent different types of information content attributes; The columns of the community feature matrix are defined as time windows to represent the time series evolution of community information status; Each element in the community feature matrix is ​​a probability estimate of the occurrence or spread of its corresponding feature in the community within the corresponding time window.

3. The network information dissemination method as described in claim 2, characterized in that: By extracting keywords, classifying topics, and embedding representations from the information to be disseminated, a normalized feature vector aligned with the row dimensions of the community feature matrix is ​​generated. From the community feature matrix, select the column vector of the current time window as the target vector; The cosine similarity between the information feature vector and the target column vector is calculated as a similarity metric.

4. The network information dissemination method as described in claim 3, characterized in that: The similarity is used to calculate the fusion weight, and the fusion weight is used to interpolate the information feature vector and the current time column vector of the community feature matrix to obtain the updated current time column vector. Replace the corresponding columns in the original community feature matrix with the updated column vectors, and update the community feature matrix.

5. The network information dissemination method as described in claim 1, characterized in that: The forwarding utility model takes the fused feature vector as input, forwards information to different neighbor nodes of the information relay point, and predicts the expected information coverage increment and propagation efficiency gain in the target community within a preset time window in the future, with the expected information coverage increment and the propagation efficiency gain as the output targets. The initial training of the forwarding utility model is based on a historical propagation record dataset, which contains feature inputs and actual propagation effect sample pairs. The forwarding utility model is established by learning the nonlinear relationships in historical data.

6. The network information dissemination method as described in claim 1, characterized in that: The step of calculating the difference between the actual coverage utility matrix and the expected coverage utility matrix, and updating the forwarding utility model and community feature matrix in conjunction with the feedback matrix, includes: The difference between the actual coverage utility matrix and the expected coverage utility matrix is ​​calculated to obtain the model prediction error matrix; The prediction error matrix and the resident demand feedback vector are used together as training data, and the parameters of the forwarding utility model are updated through online learning. The changes in residents' interests reflected by the resident demand feedback vector are incrementally integrated into the community feature matrix of the target community to update the estimate of the information feature probability distribution.

7. A network information dissemination system, characterized in that... Includes a network information dissemination method as described in any one of claims 1-6; wherein: The community feature management module is used to set up a community feature matrix for each community. When information to be disseminated is received, the module extracts the feature vector of the information to be disseminated, calculates the similarity between the feature vector and the column vector of the target community's community feature matrix, and integrates the feature vector into the current time column of the community feature matrix to obtain an updated community feature matrix. The module also calculates the change in entropy of the community feature matrix before and after the update, which is used as the dissemination value weight of the information to be disseminated to the target community. The forwarding decision module, connected to the community feature management module, is used to obtain the community spatiotemporal context features at the current moment, and input the propagation value weight and the community spatiotemporal context features into the forwarding utility model to predict the expected coverage utility generated by forwarding the information to each neighbor node within the target community; The feedback optimization module is connected to the community feature management module and the forwarding decision module, respectively. It is used to collect the feedback matrix, resident demand feedback vector and actual coverage utility matrix constructed by resident feedback, calculate the difference between the actual coverage utility matrix and the expected coverage utility matrix, and update the forwarding utility model and the community feature matrix in combination with the feedback matrix and demand feedback vector.

Images