Detecting opinion spams through supervised boosting approach

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Mining

Background:

• Product reviews are crucial for businesses and consumers, but opinion spam disrupts trust and accuracy in platforms like Google Playstore.
• Existing spam detection methods, often using time series and neural networks, are computationally intensive and limited by feature granularity.
• The need for more accurate and efficient opinion spam detection is evident, especially in multilingual mobile application marketplaces.

Purpose of the Study:

• To propose and evaluate statistical-based features for improved opinion spam detection in mobile application reviews.
• To compare the effectiveness of supervised boosting models, specifically Extreme Gradient Boost (XGBoost) and Generalized Boosted Regression Model (GBM), for spam detection.
• To assess model performance on multilingual datasets (English and Malay) to understand language-specific efficacy.

Main Methods:

• Utilized statistical features modeled through supervised boosting algorithms: Extreme Gradient Boost (XGBoost) and Generalized Boosted Regression Model (GBM).
• Evaluated the proposed methods on two multilingual datasets, comprising English and Malay mobile application reviews.
• Conducted comparative analysis to determine the most suitable model for each language dataset.

Main Results:

• XGBoost demonstrated superior performance for opinion spam detection in the English dataset.
• GBM Gaussian proved most effective for detecting opinion spam within the Malay dataset.
• The proposed statistical features achieved high detection accuracy rates: 87.43% for English and 86.13% for Malay.

Conclusions:

• Statistical features combined with supervised boosting models offer a robust approach to opinion spam detection.
• Model selection is crucial, with XGBoost and GBM Gaussian showing language-specific suitability.
• The study successfully improved spam detection accuracy in multilingual mobile app review contexts.