Positive Unlabeled Learning for Deceptive Reviews Detection

Abstract

Deceptive reviews detection has attracted significant attention from both business and research communities. However, due to the difficulty of human labeling needed for supervised learning, the problem remains to be highly challenging. This paper proposed a novel angle to the problem by modeling PU (positive unlabeled) learning. A semi-supervised model, called mixing population and individual property PU learning (MPIPUL), is proposed. Firstly, some reliable negative examples are identified from the unlabeled dataset. Secondly, some representative positive examples and negative examples are generated based on LDA (Latent Dirichlet Allocation). Thirdly, for the remaining unlabeled examples (we call them spy examples), which can not be explicitly identified as positive and negative, two similarity weights are assigned, by which the probability of a spy example belonging to the positive class and the negative class are displayed. Finally, spy examples and their similarity weights are incorporated into SVM (Support Vector Machine) to build an accurate classifier. Experiments on gold-standard dataset demonstrate the effectiveness of MPIPUL which outperforms the state-of-the-art baselines.

Classification of deceptive opinions using a low dimensionality representation

“…In Table 4 we can observe the indirect comparison of our results with those of (Banerjee and Chua, 2014) and (Ren et al, 2014) obtained with a 10 fold cross validation experiment, and then, with a 5 fold cross validation in order to make a fair comparison with the results of (Ott et al, 2011) and (Feng and Hirst, 2013). Note that the results are expressed in terms of the accuracy as those were published by the authors; the results correspond only to positive reviews of the Opinion Spam corpus because the authors experimented in that corpus alone.…”

“…The accuracy of the semi-supervised model is slightly lower (86.69%) than that of our approach (89%), although good enough. The authors concluded that the good performance of the semi-supervised model is due the topic information captured by the model combined with the examples and their similarity (Ren et al, 2014). Then, they could obtain an accurate SVM classifier.…”

“…10 fold cross-validation (Banerjee and Chua, 2014) 70.50% (Ren et al, 2014) 86.69% Our approach 89% 5 fold cross-validation (Ott et al, 2011) 89.8% (Feng and Hirst, 2013) 91.3% Our approach 89.8% Table 4: Indirect comparison of the performance. Deceptive opinions detection for positive reviews of Opinion Spam corpus (800 opinions).…”

Detecting opinion spams through supervised boosting approach

“…Prior to that, researchers such as Jindal and Liu [ 13 ] implemented a machine learning approach by adopting the logistic regression to detect opinion spams, so did Li, Ott [ 14 ], Lin, Zhu [ 15 ] and Ren and Ji [ 16 ] who likewise, used supervised machine learning to detect opinion spams. However, earlier studies by Ren, Ji [ 17 ] and Li, Chen [ 18 ] implemented a hybrid approach that combined supervised and semi-supervised machine learning to detect opinion spams. Despite this being so, there were several limitations to the machine learning approach, for example, using too many features, providing less accurate outcomes, poor flexibility and high consumption of computational time.…”

Relaxed softmax for PU learning

“…Contrary to the positive-only case, Positive and Unlabeled (PU) learning [8,17,18,27,29], leverages some non-labeled data to further improve the performance of the model in classification tasks. One of the classical methods is to identify possible negatives in this unlabeled set.…”