Research On Opinion Spam Detection Based On Deep Learning

With the continuous development of e-commerce,much more customers rate,evaluate and search for products online.Product reviews have become an important reference for most customers when they purchase products.So,the webpages with user comments have become a spot for spamming.As a result,some businesses hire writers to write fake reviews for their own products or write malicious comments to opponents,they use improper ways to profit themselves will bring a lot of adverse effects to the entire consumer industry.According to the research,it is difficult for human readers to judge the difference between the real comment and the fake comment.Therefore,it is very important to use machine learning to identify the spam reviews.The concept of deep learning is a method proposed by Geoffrey E.Hinton in 2006 that simulates the behavior of biological neural networks and processes distributed parallel information.Deep learning simulates the neuron's judgment method and transform the linear structure into the nonlinear structure.Deep learning deepens the number of traditional artificial neural network layers and adopts a variety of pre-training methods to obtain initial values of network parameters,avoiding the divergence caused by random value initialization.Convolution neural network is a kind of deep learning method,which was first proposed by Yann Lecun and applied to handwritten numeral recognition.It has achieved great success in the fields of speech recognition and image recognition.At present,the application of natural language processing has got a lot of results.This paper constructs a model of convolutional neural network(MFCNN)for spam review detection based on the standard dataset published by Ott and Li et al,which use some convolution filters with different width to represent the sentence and document.Then,this paper proposes a new activation function based on Re LU and Softplus,which named SRLU,to reduce the offset value and the gradient disappearance.In the end,this paper classifies the spam reviews by MFCNN+SRLU model,compares the result with many machine learning methods and activation function,and gets the more accurate results.This paper first extracts the word vector model through the pre-trained word vector model provided by Google News,then builds a deceptive spam detection model based on MFCNN+SRLU model,verifies the influence of the number and width of convolution filters,learning rate.By comparing with other machine learning algorithms and activation functions,this paper proves the effectiveness of the proposed method.The experimental results show that the accuracy of the method is 92.50%,the precision is 94.37%,the recall is 89.33% and the F1-score is 91.78%.Compared with other methods,the proposed MFCNN+SRLU model can effectively detect spam reviews.