CN-121684066-B - Anti-addiction early warning system and early warning method based on causality discovery network

CN121684066BCN 121684066 BCN121684066 BCN 121684066BCN-121684066-B

Abstract

The application is suitable for the technical field of artificial intelligence and the technical field of big data, and provides an anti-addiction early warning system and an early warning method based on a causality discovery network, wherein the anti-addiction early warning system comprises a central server and a plurality of client devices, and the central server is used for sending a global model to the plurality of client devices; the central server is used for receiving the local model updating quantity uploaded by each client device, adopting an aggregation formula to aggregate the local model updating quantity of each client device, generating an early warning model, pushing the early warning model to target client devices in a plurality of client devices, generating an attribution report of the current user by the target client devices through a causal discovery network in the early warning model, and generating early warning information and intervention advice of the current user in the aspect of network addiction based on the attribution report and the agent of the current user. The application is beneficial to improving the reliability of early warning information and intervention advice.

Inventors

WANG FUQIANG
LI ZONGYU
ZHANG WEIWEI
LI YANG

Assignees

湖南红普创新科技发展有限公司

Dates

Publication Date: 20260512
Application Date: 20260210

Claims (9)

1. An anti-addiction early warning system based on a causal discovery network is characterized by comprising a central server and a plurality of client devices, wherein the central server stores a global model of a federal learning architecture; the central server is used for sending the global model to a plurality of client devices; Each client device is used for receiving the global model, training the global model by using a local training sample to obtain parameters of the local model, selecting a difference value between the parameters of the local model and a parameter set of the global model as a local model updating amount, and uploading the local model updating amount to a central server; The central server is used for receiving the local model updating quantity uploaded by each client device, adopting an aggregation formula to aggregate the local model updating quantity of each client device, generating an early warning model, and pushing the early warning model to a target client device in the plurality of client devices; The target client device receives the early warning model, deploys the early warning model, acquires the behavior data of the current user from the operation log, acquires the psychological data of the current user from the psychological assessment table, and acquires the physiological data of the current user from the acquired data of the wearable intelligent device connected with the target client device; The target client device adopts a preset feature extraction model to respectively extract the behavior data of the current user, the psychological data of the current user and the physiological data of the current user, generate the features of the behavior data of the current user, the psychological data of the current user and the physiological data of the current user respectively, splice the features of the behavior data of the current user, the psychological data of the current user and the physiological data of the current user to obtain the current fusion features of the current user, input the current fusion features of the current user into an early warning model, generate the causal relationship of the network addiction behavior of the current user through a causal discovery network in the early warning model, input the causal relationship of the network addiction behavior of the current user into a counter fact reasoning module in the early warning model, generate intervention parameters of the network addiction behavior of the current user through the counter fact reasoning module, input the intervention parameters of the network addiction behavior of the current user into a generator in the early warning model, and generate an attribution report of the current user through the generator; The target client device acquires the reason of the network addiction behavior of the current user and the network addiction level of the current user from the attribution report of the current user, generates an early warning message of the current user in the aspect of network addiction when the network addiction level of the current user is greater than a preset level, inputs the reason of the network addiction behavior of the current user into the trained agent, and generates an intervention suggestion of the current user in the aspect of network addiction through the trained agent.
2. The anti-addiction early-warning system of claim 1, wherein the central server is configured to receive the local model update amount uploaded by each client device, aggregate the local model update amount of each client device using an aggregation formula, generate an early-warning model, and push the early-warning model to a target client device of the plurality of client devices, and the method comprises: the central server is used for receiving the local model updating quantity uploaded by each client device, adopting an aggregation formula to aggregate the local model updating quantity of each client device to obtain the updating parameters of the global model, acquiring the total loss value of the global model through the total loss function, determining that the global model meets the convergence condition when the total loss value is smaller than a preset value, importing the updating parameters of the global model into the framework of the global model, generating an early warning model, and pushing the early warning model to the target client devices in the plurality of client devices.
3. The anti-addiction early warning system of claim 1, wherein said current user's behavioral data comprises one or more combinations of: The method comprises the steps of using an event sequence of an application, starting a time stamp of the application, closing the time stamp of the application, running time of the application, sliding speed of a screen, typing speed of the screen, website type in a network access log, access time period in the network access log and search keywords in the network access log, wherein the application event sequence is a set of operation events of the application recorded by target client equipment according to time sequence.
4. The anti-addiction alert system of claim 1, wherein the psychological data of the current user comprises one or more of the following combinations: A score of a negative emotional state of the current user, a score of a stress level of the current user, a score of an autism of the current user; wherein, the higher the score of the negative emotion state of the current user, the score of the stress level of the current user and the score of the autism of the current user, the higher the network addiction level of the current user; wherein the lower the score of the negative emotional state of the current user, the score of the stress level of the current user and the score of the autism of the current user, the lower the network addiction level of the current user.
5. The anti-addiction alert system of claim 1, wherein the current user's physiological data comprises one or more of the following combinations: the skin conductance reaction value of the current user, the resting heart rate of the current user and the sleeping time of the current user; The higher the skin conductance reaction value of the current user and the resting heart rate of the current user are, the higher the network addiction level of the current user is, the lower the skin conductance reaction value of the current user and the resting heart rate of the current user are, and the lower the network addiction level of the current user is; The sleep time of the current user is longer, and the network addiction level of the current user is lower.
6. The anti-addiction alarm system of claim 1, The training sample comprises multimode data of a preset user and attribution reports of the preset user, wherein the multimode data of the preset user comprises behavior data, psychological data of the preset user and physiological data of the preset user; The target client device is a client device that logs in to the account of the minor.
7. The anti-addiction early warning system of claim 1, wherein the aggregate formula is defined as follows: ); Wherein, the The update parameters of the global model are expressed in the first place The parameter set of the global model is used for the next round of training; Represent the first Local model update amount uploaded by each client; Is shown in the first During round training, parameter sets of the global model; Representing a learning rate; a serial number representing the client device; Represent the first The amount of data for the individual client devices, Representing the total data volume, which is the sum of the data volumes of all client devices.
8. The anti-addiction alert system of claim 2, wherein the total loss function is defined as follows: ; ; The larger the total loss value is, the weaker the overall performance of the global model in the aspects of classification capability and causal reasoning capability is indicated, and the smaller the total loss value is, the stronger the overall performance of the global model in the aspects of classification capability and causal reasoning capability is indicated; The cross entropy loss of the global model is represented, the larger the cross entropy loss of the global model is, the weaker the classification capability of the global model is, the smaller the cross entropy loss of the global model is, and the stronger the classification capability of the global model is; The constraint loss of the global model is represented, the greater the constraint loss of the global model is, the more serious the degree of the causal graph structure learned by the global model deviates from the directed acyclic graph is, the weaker the causal reasoning capacity is, the smaller the constraint loss of the global model is, the lighter the degree of the causal graph structure learned by the global model deviates from the directed acyclic graph is, and the stronger the causal reasoning capacity of the global model is; Representing super parameters to balance constraint loss of the global model and cross entropy loss of the global model; Representing the sum of all elements on the main diagonal of the matrix; representing an adjacency matrix, the adjacency matrix being used to represent a causal graph structure; hadamard products representing the adjacency matrix; matrix indexes corresponding to Hadamard products of adjacent matrixes; Representing the dimensions of the adjacency matrix.
9. An early warning method based on the anti-addiction early warning system of claim 1, wherein the early warning method comprises the following steps: The target client device reads preset push time and judges whether the current time is push time or not; And if the current time is the pushing time, pushing an early warning message and an intervention suggestion of the current user in the aspect of network addiction to the monitoring terminal.

Description

Anti-addiction early warning system and early warning method based on causality discovery network Technical Field The application belongs to the technical field of artificial intelligence and the technical field of big data, and particularly relates to an anti-addiction early warning system and an anti-addiction early warning method based on a causality discovery network. Background With the popularization of the internet and intelligent devices, the problem of network addiction has become a global social issue, and network addiction refers to the behavior that users cannot autonomously control the use duration and frequency of a network and depend on network services excessively, so that physical and mental health, social relations, work and study and daily life of the users are negatively influenced. In order to cope with the problem, the existing network anti-addiction system is developed, but the existing network anti-addiction system adopts static rules to implement intervention, for example, a method of setting the upper limit of the daily use time period and forcing the network to be disconnected in a fixed period is adopted to limit the user behavior. Therefore, how to generate early warning information and intervention advice of the current user in the aspect of network addiction is a technical problem to be solved urgently. Disclosure of Invention The embodiment of the application aims to provide an anti-addiction early warning system based on a causal discovery network, which aims to solve the technical problem of how to generate early warning messages and intervention suggestions of a current user in the aspect of network addiction. In a first aspect, an embodiment of the present application provides an anti-addiction early warning system based on a causal discovery network, where the anti-addiction early warning system includes a central server and a plurality of client devices, where the central server stores a global model of a federal learning architecture; the central server is used for sending the global model to a plurality of client devices; Each client device is used for receiving the global model, training the global model by using a local training sample to obtain parameters of the local model, selecting a difference value between the parameters of the local model and a parameter set of the global model as a local model updating amount, and uploading the local model updating amount to a central server; The central server is used for receiving the local model updating quantity uploaded by each client device, adopting an aggregation formula to aggregate the local model updating quantity of each client device, generating an early warning model, and pushing the early warning model to a target client device in the plurality of client devices; The target client device generates an attribution report of the current user through a causal discovery network in the early warning model based on a predefined mode, and generates early warning messages and intervention suggestions of the current user in terms of network addiction based on the attribution report of the current user and an agent. In a possible implementation manner of the first aspect, the central server is configured to receive a local model update amount uploaded by each client device, aggregate the local model update amount of each client device using an aggregation formula, generate an early warning model, and push the early warning model to a target client device in the plurality of client devices, where the method includes: the central server is used for receiving the local model updating quantity uploaded by each client device, adopting an aggregation formula to aggregate the local model updating quantity of each client device to obtain the updating parameters of the global model, acquiring the total loss value of the global model through the total loss function, determining that the global model meets the convergence condition when the total loss value is smaller than a preset value, importing the updating parameters of the global model into the framework of the global model, generating an early warning model, and pushing the early warning model to the target client devices in the plurality of client devices. In a possible implementation manner of the first aspect, the target client device generates, through a causal discovery network in an early warning model, an attribution report of a current user based on a predefined manner, and generates, based on the attribution report of the current user and an agent, an early warning message and an intervention advice of the current user in terms of network addiction, including: The target client device receives the early warning model, deploys the early warning model, acquires the behavior data of the current user from the operation log, acquires the psychological data of the current user from the psychological assessment table, and acquires the physiological data of the current user