Research On Constraint Based Interpretable Fraud Reviewer Detection Model

Product reviews on e-commerce platforms to some extent affect consumers’ purchasing decisions.Driven by interests,some businesses may hire fraudulent commentators to publish a large number of fraudulent reviews to promote or belittle target products,which has a certain negative impact on e-commerce platforms.To detect these fraudulent commentators,domestic and foreign scholars have conducted extensive research and achieved good detection results.However,these detection models are unable to fully utilize important information in the original dataset,which has a certain impact on detection accuracy.Although graph neural networks have improved the accuracy of model detection to a certain extent,these detection methods are mostly black box models,lacking a reasonable explanation of the detection results.At present,some studies have used post hoc interpretation methods to explain the model results,but these methods are not clear enough in explaining the reasons for model classification,which limits the application of the model.To address the above issues,this thesis proposes a constraint based model intrinsic interpretation method and a fraud commenter detection method based on a dual graph structure.Firstly,to address the issue of relatively unclear explanations obtained by existing detection models using post hoc interpretation methods,a constraint based model intrinsic interpretation method is proposed.This method introduces the definition of explanatory constraints and extracts key structures,namely meta graphs,from heterogeneous information networks constructed from e-commerce platform datasets;Using meta graphs as clues for graph traversal and guiding Monte Carlo tree search with heterogeneous information networks as root nodes;By comparing each node vector in the search tree with the model classification vector,the importance score of nodes in a heterogeneous information network is obtained.The set formed by nodes with higher scores is the internal explanatory part of the heterogeneous information network.Secondly,in response to the problem that existing detection models cannot fully utilize the information in e-commerce platform datasets,a fraud commenter detection model with fusion interpretation mechanism is proposed.This method takes heterogeneous information networks and meta graphs as inputs,encodes the two graphs by constructing an encoder with an upper and lower layer structure,and obtains their embedding vectors;Design a prototype iteration method based on prototype learning,calculate the similarity between the embedding vectors and prototype vectors of two graphs,and update the prototype;Analyze the classification loss and prototype loss,complete model training,and input the prototype vector and the embedding vector of the heterogeneous information network into the classifier to obtain the classification results.Finally,the two algorithms proposed in this thesis and the existing algorithms are tested on Amazon dataset and Yelp dataset,and the experimental results are compared and analyzed.