Search

CN-122022888-A - Social situation network information demand prediction method and system based on time dynamics and intelligent fusion

CN122022888ACN 122022888 ACN122022888 ACN 122022888ACN-122022888-A

Abstract

Aiming at the problems of short time dynamics modeling of social situation factors, low news identification precision, static fusion mechanism and non-unified framework of multi-source factors in the existing network information demand prediction, the method and the system for modeling and intelligent fusion of the social situation factors based on time dynamics are provided. The method collects three types of situation factors in a sampling period of 1 hour, classifies and complements the situation factors according to INGWERSEN frames, determines optimal lag time through CCF traversal [24,24] hours, combines the Granges causal test (p < 0.01) to provide time dynamics characteristics, identifies sensitive news by using a BiGRU +CRF classifier and quantifies emotion intensity, establishes a dynamic weight mechanism according to factor types and differential attenuation functions, and fuses the output characteristic vectors. The method can be used as an independent module for a time sequence prediction model, wherein the MSE of an emergency scene in a BAI data set is reduced by 47.8%, the news identification accuracy is 91.2%, the community scene prediction accuracy is increased by 32.1%, the operation and maintenance response time is shortened by 57%, and the prediction accuracy and the robustness are effectively improved.

Inventors

  • SUN ZHIXIN
  • ZHAO XIAOYU
  • XU YUHUA

Assignees

  • 南京邮电大学

Dates

Publication Date
20260512
Application Date
20260129

Claims (10)

  1. 1. A social situation network information demand prediction method based on fusion of time dynamics and intelligence is characterized by comprising the following steps: the method comprises the steps of step 1, collecting and classifying social situation factors, collecting physical environment situation factors, social organization situation factors and time structure situation factors by taking 1 hour as a sampling period, supplementing data with the loss rate less than or equal to 2% through linear interpolation, writing the factors into a classification table according to a INGWERSEN three-layer situation classification frame and combining with a rule set, and outputting structured data containing fields { timestamp, factor_type, value }, wherein the rule set comprises weather, physical environment, news, social organization, holiday and time structure; step 2, identifying time dynamics characteristics of the situation factors, traversing the cross correlation function CCF of each factor sequence obtained in the step 1 and the target information demand sequence, obtaining CCF (k) by taking 1 hour as step length within [24,24] hours after k, and selecting the maximum value corresponding to the CCF (k) As an optimal lag time, to Recording for the center that the correlation coefficient falls to half the peak value and to the threshold value Form a time-dynamic characteristic parameter , wherein, Represents the first The time at which the class signal reaches its peak, and is defined herein as ; Represents the first The corresponding time when the value of the class signal is reduced from the peak value to half of the peak value; represents the first The value of the class signal decreases from the peak value to the threshold value Verifying the unidirectional causality through the Granges causality test, and eliminating the factors without causality or with the opposite causality; Step 3, intelligent news sensitivity recognition and emotion quantification, namely inputting a BiGRU +CRF classifier after dividing and deactivating news text in social organization situation factors, and outputting labels { disaster/policy/folk/non-sensitive } and emotion intensity scores And writing the result to the structured record ; Step 4, dynamic weight adjustment based on time attenuation function according to Selecting a corresponding decay function Class-level weights are calculated , wherein, As the basis weight for the category, For and context factor category The corresponding differential attenuation function is used to determine, 、 、 Respectively, by time dynamics characteristic parameter vector Determining parameters of peak time, attenuation constant and attenuation type, adopting asymmetric exponential attenuation alpha >0 for social tissue situation factors, adopting symmetric exponential attenuation alpha=0 for physical environment factors and adopting step function for time structure factors, and executing normalization on all situation categories at the same time ; Step 5, intelligent fusion of multiple source situation factors, time alignment of the factors according to k, using Manipulation translates the factor sequence backwards Hours: Post-alignment factor features And the step 4 is carried out Common input attention module, calculate And outputs the fused context feature vector 。
  2. 2. The method according to claim 1, wherein in the step 2, the calculation formula of the cross correlation function CCF is: wherein k is the number of hysteresis steps, For the value of the context factor at time t, Selecting by systematically traversing different lag times k for the information demand to take the value at time t+k Taking k reaching the maximum value as the optimal lag time, further taking the optimal lag time as the center, calculating the time interval when the correlation coefficient is reduced to half of the peak value and to the preset threshold epsilon, thereby obtaining the time dynamics characteristic parameter vector of each type of situation factors 。
  3. 3. The method according to claim 1, wherein in step 2, the graininess causal test is used to verify causal effect of the context factor on the information demand, excluding the inverse causal relationship, the test statistic is F statistic, significant graininess causal relationship is considered to exist when p <0.01, and the hysteresis order selection principle is AIC minimum and no more than 8 th order.
  4. 4. The method of claim 1, wherein the BiGRU +CRF news sensitivity classifier architecture in step 3 includes a word embedding layer, a bi-directional GRU encoding layer, a full connection layer, and a CRF decoding layer, and the classifier outputs a type tag and an emotional intensity score of the sensitive news.
  5. 5. The method according to claim 1, wherein in the step 4, the dynamic weight calculation formula is: Wherein, the For the moment of time Corresponding to the category of the situation factor Is used for the normalization of the fusion weights of (a), As the basis weight for the category, For and context factor category The corresponding differential attenuation function is used to determine, 、 、 Respectively, by time dynamics characteristic parameter vector Determined peak time, decay constant and decay type parameters, denominator pairs Traversing INGWERSEN three layers of context categories to realize weight normalization, selecting corresponding decay function types according to the context factor types, namely asymmetric decay functions Adopts an exponential decay form to realize the time dynamics characteristics of quick activation and slow decay, and a symmetrical decay function Realizing symmetrical activation and attenuation, step function Realizing a step-type continuous mode; , Wherein, the Is the reference time constant of the original signal, The coefficients are modified for the time constant of the original signal, Is the time constant of the envelope of the signal, For the time interval in which the signal value falls from the peak value to the threshold value, Representing the time difference between the current time and the peak time of the original signal, Representing the current time of day, Representing the moment at which the original signal reaches its peak.
  6. 6. The method of claim 1, wherein in step 5, the multi-source context factor fusion uses an attention mechanism to calculate the attention weight of each context factor, and the adaptive fusion of context factors is realized Comprising a layer 2 fully connected network, the output of which is obtained by softmax 。
  7. 7. The method of claim 1, wherein the input of the BiGRU +crf classifier in step 3 uses a 300-dimensional word vector, the number of bi-directional GRU layers is 2, the concealment dimension is 128, the loss function is a weighted sum of cross entropy and CRF loss, and the training batch size is 64.
  8. 8. A social context network information demand prediction system based on a fusion of time dynamics and intelligence for implementing the method of any one of claims 17, the system comprising: the situation factor acquisition module is used for acquiring three types of multi-source social situation factor data including physical environment situation, social organization situation and time structure situation; the situation factor classification module is used for automatically classifying the situation factors according to INGWERSEN three-layer situation classification frameworks; the time dynamics feature identification module is used for identifying the optimal lag time and time dynamics features of the situation factors through a cross correlation function and a Grangel causal test; The news sensitivity recognition module is used for recognizing sensitive news events through a BiGRU +CRF classifier and quantifying emotion intensity; The dynamic weight adjustment module is used for dynamically adjusting the fusion weight of each situation factor according to the time attenuation function; the multi-source factor fusion module is used for uniformly fusing the situation factors subjected to time alignment and weight adjustment; And the prediction output module is used for outputting the fused situation characteristic vector for the prediction model.
  9. 9. The system of claim 8, wherein the temporal dynamics characterization module is further configured to calculate a peak lag time, decay half-life, and effect duration of the contextual factor and generate a table of temporal dynamics characterization parameters.
  10. 10. The system of claim 8, wherein the dynamic weight adjustment module is further configured to select a corresponding decay function type according to a type of the context factor to implement the differential weight adjustment.

Description

Social situation network information demand prediction method and system based on time dynamics and intelligent fusion Technical Field The invention relates to the technical field of artificial intelligence and network intelligent operation, and relates to the field of network flow prediction, in particular to a social situation network information demand prediction method and system based on time dynamics and intelligent fusion, which aim at the problems of time lag effect modeling, time dynamics feature recognition, dynamic weight fusion based on a time decay function and intelligent multi-source situation factor integration of social situation factors. Background With the rapid development of the internet and mobile communication technology, network information demand prediction has become a core technology for network resource optimal configuration and intelligent operation. The prior art has recognized the important impact of social context factors (such as weather, holidays, news events, etc.) on network information requirements, but the following key technical drawbacks still exist in terms of the time dynamics modeling and intelligent fusion of the social context factors: It should be noted that, in the prior art, methods for predicting network traffic based on BiGRU, ARIMAX and other hybrid models exist, and the methods improve prediction accuracy to a certain extent by integrating a statistical model and a deep learning model. However, these methods still have the disadvantage of time dynamics modeling of social context factors that, although some of them simply mention time alignment of social context factors (e.g., several hours of weather lag, several hours of news lag), they do not systematically identify and model the time dynamics characteristics of the context factors (e.g., peak lag time, decay half-life, effect duration), nor establish a dynamic weight adjustment mechanism based on a time decay function, resulting in inadequate prediction accuracy in emergency scenarios. The invention focuses on the temporal dynamic feature recognition and intelligent fusion of the social situation factors, can be used as an independent social situation factor processing module, can be used as exogenous input of any time sequence prediction model, can also be used in combination with the existing mixed prediction model, and improves the prediction precision. First, social context factors lack time dynamics modeling. In the prior art, social situation factors are taken as static exogenous variables to be directly input into a prediction model, and the time lag effect and dynamic evolution characteristics of the situation factors affecting information requirements are ignored. In fact, there is a significant time difference in the impact of different environmental factors on information demand, namely the impact of weather changes on information demand typically lags by 4, 6 hours (e.g., online entertainment demand increases after rainfall), the impact of sensitive news events may peak within 3 hours, while the impact of holidays takes effect almost instantaneously. It should be noted that, as classical statistical analysis methods, cross Correlation Functions (CCF) and glabellar causal tests have been widely used in the fields of economics, finance, etc., but in the field of network traffic prediction, the prior art does not systematically apply these methods for time-dynamic feature recognition. In the prior art, the lag time is determined by adopting an empirical value or simple correlation analysis, the optimal lag time is not systematically identified by CCF traversal analysis (24 to 24 hours), and the causal direction is not verified by the Grangel causal test, so that complete time dynamics characteristic parameters such as the peak lag time, decay half-life, effect duration and the like of the situation factors cannot be accurately identified. The technical result is that the MSE may be slightly improved (reduced by 5% 10%) in normal situations, but the prediction fails (the error may exceed 23 times of the normal value) in emergency situations, and the peak effect of the emergency cannot be accurately captured. Second, news event recognition accuracy is not sufficient. In the prior art, for text social factors such as news, social media and the like, a simple keyword matching or emotion dictionary method is mostly adopted, so that the type (disaster/policy/folk) and intensity of sensitive news cannot be accurately identified. For example, the conventional method may misjudge local conventional events such as "traffic accident at a place, injury of 2 persons" as sensitive events, or misjudge potential hot events such as "attention caused by treatment problem of a teacher at a place". The recognition error causes that the prediction model cannot accurately capture demand fluctuation of news drive, and the prediction error is obviously increased under the traffic surge scene caused by sensitive news.