Online Reputation systems, such as eBay, Amazon, Alibaba, etc., are a novel and active part of E-commerce environments. These corporations use reputation reporting systems for trust evaluation by measuring the overall feedback ratings given by buyers, which enables them to compute the reputation score of their products. Such evaluation and computation processes are closely related to sentiment analysis and opinion mining. These techniques incorporate new features into traditional tasks, like polarity detection for positive or negative reviews. For instance, lack of honesty or effort in providing the feedback reviews, by which users might create phantom feedback from fake reviews in order to support their reputation. The “all excellent reputation” problem is common in the e-commerce domain, where most of the feedback ratings are positive, leading to high reputation scores for the sellers. Another problem is that sellers can write unfair reviews to endorse or reject any given targeted product since a higher reputation leads to higher profits. For reviews to reflect genuine user experiences and opinions, such unfair reviews must be detected.
The problem of unfair reviews may be aggravated by the collusion of multiple users where the fear of bad ratings causes reviewers to collude, manipulate and deceive others. It is often the case that unfair ratings have a different statistical pattern than the fair ratings. The main objective of this study is to offer a novel and comprehensive solution for designing a new model to obtain the most accurate reputation system, which addresses the existing issues, such as fake feedback reviews and unfair reviews from opinion reviews, collusion and manipulation, as well as the ”all good reputation” issue that is being currently encountered by reputation systems.
The purpose of the proposed research is to use a statistical technique for excluding unfair ratings and to illustrate its effectiveness through simulations. In order to do that, we first started by analysing online movie reviews using Sentiment Analysis (SA) methods in order to detect fake reviews. SA and text classification methods were applied to datasets of movie reviews. More specifically, we compared five supervised machine learning algorithms : Naïve Bayes (NB), Support Vector Machine (SVM), K-Nearest Neighbours (KNN-IBK), KStar (K*) and Decision Tree (DT-J48) for sentiment classification of reviews, using three different datasets. In order to evaluate the performance of sentiment classification, this work has implemented accuracy, precision, recall and F-measure as performance measures. The measured results of our experiments show that the SVM algorithm outperforms other algorithms, and that it reaches the highest accuracy not only in text classification, but also in detecting fake reviews. Second, we carried out comparison study of four supervised machine learning algorithms : Naïve Bayes (NB), Decision Tree (DT-J48), Logistic Regression (LR) and Support Vector Machine (SVM) for sentiment classification using three datasets of Amazon reviews, including Clothing, Shoes and Jewelry reviews, Baby reviews as well as Pet Supplies reviews. In order to evaluate the performance of sentiment classification, this work has implemented accuracy, precision and recall as performance measures. Our experiments’ results show that the Logistic Regression (LR) algorithm is the best classifier with the highest accuracy as compared to the other three classifiers, not merely in text classification, but in unfair reviews detection as well. In addition, we have calculated reputation scores from users’ feedback based on a Sentiment Analysis Model (SAM), in order to obtain useful information from reviews, based on a Logistic Regression algorithm with two different feature selections. Experimental results based on two different datasets demonstrate the effectiveness of our approach in capturing reputation information from reviews.
| Date | 3 Jul 2020 |
|---|
| Original language | French |
|---|
| Awarding Institution | - École de technologie supérieure
|
|---|
| Supervisor | Abdelouahed Gherbi (Supervisor) |
|---|