Fraudulent user detection model construction method based on graph attention network

Abstract

The invention discloses a fraudulent user detection model construction method based on a graph attention network, and the method comprises the steps: firstly importing user information with a label, and separating the text content of a user; preprocessing the user information, and representing user behavior characteristics and text characteristics by using a plurality of characteristics; secondly,importing text content of a user, and calculating similarity between sentence embedding of user comments after the text content is preprocessed; constructing a user network by taking the user information features as vertexes and the comment relevancy as edges; and finally, learning the user network by using a graph convolution network, adaptively aggregating neighbor information in the graph convolution network by using an attention mechanism, and obtaining a model after repeated training. According to the method, the graph attention network is used for training on the user-user network, theglobal position information of the user in the network is deeply mined, neighbor weights are distributed in a self-adaptive mode, and therefore the accuracy and stability of model detection are improved.

Description

technical field [0001] The invention relates to a method for constructing a fraudulent user detection model based on a graph attention network, and belongs to the intersecting technical fields of fraud detection, natural language processing, graph neural network, and deep learning. Background technique [0002] The openness of the Internet and the monetary rewards of crowdsourcing tasks have stimulated a large number of fake users to write fake reviews and publish advertisements to interfere with users' judgment. Fraudulent users are characterized by subjectivity and diversification. Therefore, manual identification is difficult and costly, but the accuracy rate is not ideal. Therefore, there are two popular solutions in the industry, which are rule-based and deep learning-based. [0003] The rule-based method is mainly to classify users by analyzing and learning the behavior characteristics of users, and obtaining the classification rules of the characteristics. The method...

AI Technical Summary

Problems solved by technology

However, there are the following shortcomings: (1) It is impossible to extract the correlation information and global information between users, that is, the structural information cannot be extracted (2) Noise problem, large-scale text information often contains a lot of invalid information (3) The pooling strategy cannot Automatic selection of user feature combinations with contributions (4) does not effectively use user behavior features

In this way, the structural information between users can be extracted and the global information can be learned to solve the shortcomings (1) in the traditional model, but it cannot solve the shortcomings (2)(3)(4) in the traditional model, and because the graph is no weight undirected graph, so it is impossible to further extract structural information

Images