Detecting opinion spams through supervised boosting approach

Abstract

Product reviews are the individual’s opinions, judgement or belief about a certain product or service provided by certain companies. Such reviews serve as guides for these companies to plan and monitor their business ventures in terms of increasing productivity or enhancing their product/service qualities. Product reviews can also increase business profits by convincing future customers about the products which they have interest in. In the mobile application marketplace such as Google Playstore, reviews and star ratings are used as indicators of the application quality. However, among all these reviews, hereby also known as opinions, spams also exist, to disrupt the online business balance. Previous studies used the time series and neural network approach (which require a lot of computational power) to detect these opinion spams. However, the detection performance can be restricted in terms of accuracy because the approach focusses on basic, discrete and document level features only thereby, projecting little statistical relationships. Aiming to improve the detection of opinion spams in mobile application marketplace, this study proposes using statistical based features that are modelled through the supervised boosting approach such as the Extreme Gradient Boost (XGBoost) and the Generalized Boosted Regression Model (GBM) to evaluate two multilingual datasets (i.e. English and Malay language). From the evaluation done, it was found that the XGBoost is most suitable for detecting opinion spams in the English dataset while the GBM Gaussian is most suitable for the Malay dataset. The comparative analysis also indicates that the implementation of the proposed statistical based features had achieved a detection accuracy rate of 87.43 per cent on the English dataset and 86.13 per cent on the Malay dataset.

Spam Review Detection Techniques: A Systematic Literature Review

“…Previous works show that it is not practical to implement spam review detection methods without training [83]. The learner's accuracy is much higher when trained on real-world datasets [84]. Table 15 shows that most of the existing studies used precision, f-measure, accuracy and ranked normalized discounted cumulative gain parameters to evaluate the performance of spam review detection methods.…”

Phishing detection system using nachine learning classifiers

“…It is a popular machine learning software program developed in Java at Waikato University, New Zealand [23,24]. WEKA supports a number of standard data mining tasks, including data pre-processing, clustering, classification, regression, visualization and selection of features [4,19,25]. Feature selection methods were used to identify and remove irrelevant and redundant attributes from data that do not contribute to the accuracy of a predictive model [26].The features of the phishing website were first trained and then classified by using significant features.…”

“…The user will get fraud by entering their confidential information such as password, bank details and account credentials into the fake websites [1][2][3]. The criminal will then use the information provided to access the account to buy stuff, transfer money, or other damaging activities [3,4]. For example, in 2016 the phishing attack up to 65% worldwide which costs about $1.6 million [5].…”

Challenges and Development in Malay Natural Language Processing

“…Different morphological usage also hindered the adoption of English text processing techniques into Malay [2]. The various accents from different parts of Malaysia along with the informal abbreviation also contributed to the lack of congruency in the language (Hazim et al [8]). [3]sought to create a new electronic Malay computational lexicon to facilitate Malay morphological research and the updating of existing Malay dictionary.…”

“…The work reported in [8] explored the modelling of supervised boosting approaches along with statistical-based features to recognize opinion spams in the mobile application marketplace. The authors utilized the readily available Yelp's hotel review as English dataset while compiled the reviews of randomly chosen Malaysian applications from Google Playstore.…”