Spam and scam detection through text analysis

Abstract

This report summarizes an experimental study to detect spammer and scammer existence in e-commerce platform. The combination studies of analysing business review and rating were used to categorize the text review into two classifications, namely Truthful and Deceptive in which were classified further into Positive and Negative classes. Background knowledge for manual data labeling is discussed later. In this study, a sub-domain of Machine Learning Processing, such as Natural Language Processing (NLP) was implemented for the machine to simulate and classify the given text in human ability degree. The raw corpus collection was predicted with the application of TFIDF Transformer with Count Vectorizer initialization. Furthermore, attention mechanism was believed to pay greater attention to certain factors and help addressing the text focus during the data processing. Hence, the application of attention mechanism may enhance the output prediction accuracy and Transformer model was also considered in this study. The experimental model comparison was made between the integration of a single and multiple classifiers in BERT model. Some programming modules, such as, PyTorch, Scikit-Learn, Keras, spaCy and Natural Language Toolkit (NLTK) were widely used in this experiment.