Classification of deceptive opinions using a low dimensionality representation

Abstract

Opinions in social media play such an important role for customers and companies that there is a growing tendency to post fake reviews in order to change purchase decisions and opinions. In this paper we propose the use of different features for a low dimension representation of opinions. We evaluate our proposal incorporating the features to a Support Vector Machines classifier and we use an available corpus with reviews of hotels in Chicago. We perform comparisons with previous works and we conclude that using our proposed features it is possible to obtain competitive results with a small amount of features for representing the data. Finally, we also investigate if the use of emotions can help to discriminate between truthful and deceptive opinions as previous works show to happen for deception detection in text in general.

Detecting Deceptive Opinions: Intra and Cross-Domain Classification Using an Efficient Representation

“…We only show the results obtained with SVM because its performance was the best. 5 For all the experiments we have performed a 10 fold cross-validation procedure in order to study the effectiveness of the SVM classifier with the different representations. For simplicity, we have used LibSVM e which implements a C-SVC version of SVM with a radial basis function.…”

Fake Reviews Detection: A Survey

“…The proposed model was evaluated on two datasets (AMT and Deception dataset). The results showed that the proposed model achieved the best results with 82.9% accuracy on AMT datasets compared to the state of art methods such as LIWC and unigram with SVM [77], LIWC feature and four n-grams with SVM [140], recurrent convolutional neural network [160], profile alignment compatibility method [161], sparse additive generative model [100], lexicalized production rules with SVM [123], convolutional neural network and gated recurrent neural network [154]. Furthermore, the proposed model performed well with 80.8% accuracy on deceptive dataset compared to the state of art method [77].…”

“…They used a support vector machine and Naïve Bayes as classification algorithms. The proposed model was evaluated on a dataset consists of 'Death penalty', 'Abortion' and 'Best Friend' domains [140]. The experimental results showed that the proposed model performed better than SVM with LIWC, LDA& words, and Deep syntax & words [138] in identifying fake reviews.…”

A survey on author profiling, deception, and irony detection for the Arabic language

“…With those results, the authors conclude that PU‐learning show to be appropriate for detecting opinion spam. Character n ‐grams in tokens, the sentiment score and LIWC linguistic features such as pronouns, articles, and verbs (present, past, and future tenses) were used in Cagnina and Rosso () for the detection of opinion spam. The best results are obtained with a Naïve Bayes classifier and the combination of character 4 ‐grams in tokens and LIWC features for the representation of the opinions.…”