Spam review detection using self attention based CNN and bi-directional LSTM

Abstract

Opinion reviews are a valuable source of information in e-commerce. Indeed, it benefits users in buying decisions and businesses to enhance their quality. However, various greedy organizations employ spammers to post biased spam reviews to gain an advantage or to degrade the reputation of a competitor. This results in the explosive growth of opinion spamming. Due to its nature and their increasing volume, spam reviews are a fast-growing serious issue on the internet. Until now, researchers have developed many Machine Learning (ML) based methods to identify opinion spam reviews. However, the traditional ML methods cannot effectively detect spam messages due to the limited feature representations and the data manipulations done by spammers to escape from the detection mechanism. As an alternative to ML-based detection, in this paper, we proposed a Deep Learning (DL) based novel framework called Self Attention-based CNN Bi-LSTM (ACB) model to learn document level representation for identifying the spam reviews. Our approach computes the weightage of each word present in the sentence and identifies the spamming clues exists in the document with an attention mechanism. Then the model learns sentence representation by using Convolution Neural Network (CNN) and extracts the higher-level n-gram features. Then finally, sentence vectors are combined using Bi-directional LSTM (Bi-LSTM) as document feature vectors and identify the spam reviews with contextual information. The evaluated experiment results are compared with its variants and the result shows that ACB outperforms other variants in terms of classification accuracy.