半監督式深度學習假評論預測模型

Abstract

隨著線上購物日益普及，消費者越來越依賴線上消費者評論來做購買決策。然而，逐漸增長的假評論可能扭曲消費者的購買決策，並對線上市場的信譽構成巨大威脅。以往大多數的假評論偵測模型都沒有考慮到評論中隱含的面向和情感特徵，以及評論和評論者之間的互動關係，而且，它們也沒有考慮到真實評論與假評論，以及已標記評論與未標記評論分佈不平衡的問題。為了解決這些問題，本研究提出了一種新穎的半監督式深度學習模型來偵測假評論和假評論者。所提出的模型包含六個模組：評論者特徵提取、評論特徵提取、特徵互動、平行共注意力、對比式學習、正-負-未標記學習 (PNU 學習)。首先，我們使用前兩個模組來提取評論和評論者的特徵。接著，我們採用特徵互動模組來學習評論和評論者之間的互動關係，並利用平行共注意力模組來學習所提取的評論和評論者特徵之間的相互關係。然後，我們利用對比式學習模組來增強評論和評論者的表徵。最後，我們運用 PNU 學習模組從未標記資料中迭代找出可靠的正樣本和負樣本，並將它們納入已標記資料集中，以解決真實評論與假評論以及已標記評論與未標記評論分佈不平衡的問題。實驗結果顯示，本研究提出的模型在精確率 (precision)、召回率 (recall) 和 F1-分數 (F1-score) 方面均優於比較模型。我們的模型可以幫助平台提供可靠的線上消費者評論，進而提高消費者對線上評論的信任，並為企業提供公平的競爭環境。

With increasing online shopping, consumers are progressively depending on online consumer reviews (OCRs) to make their purchasing decisions. However, the growing prevalence of fake reviews distorts consumer purchase decisions and poses great threat to the credibility of online marketplaces. Most previous fake review detection models do not consider the aspect and sentiment features implicitly hidden in reviews, and the interaction relationships between reviews and reviewers. Nor do they consider the imbalanced distributions of genuine and fake and of labeled and unlabeled reviews. Therefore, in this study, we propose a novel semi-supervised deep learning model to detect fake reviews and reviewers. The proposed model contains six modules namely, reviewer feature extraction, review feature extraction, feature interaction, parallel co-attention, contrastive learning, and positive-negative-unlabeled (PNU) learning. First, we use the first two modules to extract the features of reviews and reviewers. Next, we employ the feature interaction module to learn the interaction relationships between reviews and reviewers, and the parallel co-attention module to learn inter-relationships among the extracted review and reviewer features. Also, we utilize the contrastive learning module to enhance the representations of reviews and reviewers. Last, we exploit the PNU learning module to iteratively identify reliable positive and negative samples from the unlabeled data and include them in the labeled dataset to deal with the imbalanced distributions of genuine and fake and of labeled and unlabeled reviews. The experimental results show that the proposed model outperforms the compared models in terms of precision, recall and F1-score. Our model can help platforms provide reliable online consumer reviews (OCRs), which in turn improves consumer trust in OCRs, and offers fair competition among businesses.