2026-10-07T21:44:45Z https://api.figshare.com/v2/oai

oai:figshare.com:article/32640516 2026-10-01T16:30:31Z portal_1172 item_type_8 month_year_10_2026

Intelligent methods for analyzing veracity and helpfulness of online reviews Alimuddin Melleng (24169311) PUREID: 618677002 Fake reviews detection helpfulness prediction deep learning machine learning natural language processing sentiment analysis Online reviews are pivotal in influencing consumer purchasing decisions, making the detection of fake reviews and the prediction and ranking of review helpfulness critical areas of study. This thesis presents a comprehensive exploration of various Machine Learning solutions to address these challenges, employing an array of data representations and advanced neural network models. Initially, the investigation focuses on the effectiveness of different data representations, including document embeddings, n-grams, emotion, and noun phrases, for detecting fake reviews and predicting helpfulness reviews. These representations are assessed using advanced model deep learning models such as BiLSTM, LSTM, GRU, CNN, and MLP, across four datasets: Hotel, Restaurant, Amazon, and Yelp. For the helpfulness prediction, the methodology established for fake reviews is specifically applied to the Amazon dataset. Further, the thesis proposes a novel unsupervised method for quantifying the helpfulness of reviews based on their relevance, emotional intensity, and specificity. This involves performing individual rankings for each characteristic, which are then amalgamated into a final helpfulness ranking. Empirical evaluations on Amazon product reviews validate the effectiveness of this approach against a contemporary baseline. Additionally, this thesis advocates a multi-task learning approach to simultaneously tackle fake reviews detection and review helpfulness prediction. Both tasks use the same features with fake reviews and helpfulness prediction and employ five deep learning models. By integrating these tasks, shared information among features is leveraged to enhance the performance of each task. Pre-trained RoBERTa embeddings are utilized for all data representations, and ensemble learning techniques are employed to improve prediction accuracy and reduce the risk of overfitting .The combination of different data representations, especially when fused using early and late data fusion techniques, significantly outperforms single data representation models. The findings from this research not only advance the field of NLP and machine learning but also provide practical methodologies to enhance the reliability and utility of online review platforms. These comprehensive approaches to tackling the twin challenges of fake reviews detection and review helpfulness prediction offer valuable insights and tools for both consumers and retailers in the digital marketplace. 2026-10-01T16:30:31Z Text Thesis 10.17034/32640516.v1 https://figshare.com/articles/thesis/Intelligent_methods_for_analyzing_veracity_and_helpfulness_of_online_reviews/32640516 All Rights Reserved