统计与数据科学学院知识库

电子商务的日益发展，改变了人们的日常消费习惯，网上购物成为消费主要途径。 在线评论作为消费者购物的一个重要依据，成为商家和买家关注的焦点。好评率高的商 家店铺更容易获得消费者的青睐。为提升店铺好评率，部分商家利用虚假评论误导消费 者来获取利益。监管部门对于虚假评论店铺也制定了惩罚措施，并对虚假评论进行识别， 但是商家进行虚假评论的方式更加隐蔽，很难利用人工方法识别海量评论信息。 为快速准确地识别虚假评论，本文试图建立一套虚假评论识别体系，包括：数据源 获取、文本数据清洗、训练集标注、模型选择与模型应用。 首先通过专家指导和机器学习标注相结合构建训练数据集，降低真实评论错误标注 的比例，从而提高训练数据的预测能力。其次使用半监督学习的方法，利用少量标记样 本，减少标记样本的工作量，利用 PU 学习算法与朴素贝叶斯、支持向量机、fastText、 GBDT、XGBoost、LightGBM 不同分类器进行训练，选取最优分类器与 PU 学习算法结 合。最后对预测结果进行可视化分析，对比虚假评论和真实评论之间的差异。PU 学习 算法是一种半监督学习，通过将所有正样本和未标记样本进行随机组合来创建训练集。 简化了数据标注的流程并提高了分类精度。PU 学习算法尤其适用于正例的数量有限并 且拥有大量未标记的数据情况，该算法在虚假评论识别领域得到广泛应用。 作为应用，利用网络爬虫技术采集电商平台的商品评论实例数据。通过专家指导和 机器学习方法部分标注真实评论数据，利用 PU 学习算法进行分类。实例结果表明：本 文方法具有良好的虚假评论识别的性能，这为消费者和监管部门提供了新的方法，具有 实际应用价值。

The increasing development of e-commerce has changed people's daily consumption habits, online shopping has become the main way of consumption. Online reviews, as an important basis for consumer shopping, have become the focus of attention of merchants and buyers. Merchant stores with high praise rates are more likely to win consumers' favor. In order to improve the store's positive rate, some merchants use false reviews to mislead consumers to gain benefits. Supervisory departments have also formulated punitive measures for fake review shops and identified fake reviews, but the way merchants conduct fake reviews is more concealed and it is difficult to use artificial methods to identify massive reviews. In order to quickly and accurately identify false comments, this thesis attempts to establish a set of false comment recognition methods, including: data source acquisition, text data cleaning, training set annotation, model selection and model application. Firstly, a training data set is constructed through the combination of expert guidance and machine learning annotation to reduce the proportion of false annotations of real reviews, thereby improving the prediction ability of the training data. Secondly, using semi-supervised learning method, using a small number of labeled samples to reduce the workload of labeled samples, using PU learning algorithms and Naive Bayes, support vector machine, fastText, GBDT, XGBoost, LightGBM and different classifiers to train and select the optimal classifier to combine PU learning algorithms. Finally, a visual analysis of the prediction results is performed to compare the differences between fake and real reviews. The PU learning algorithm is a semi-supervised learning that creates a training set by randomly combining all positive samples and unlabeled samples. Simplified data labeling process and improved classification accuracy. The PU learning algorithm is particularly suitable for cases with a limited number of positive examples and a large amount of unlabeled data. This algorithm is widely used in the field of false comment recognition. As an application, we use the web crawler technology to collect product review instance data of the e-commerce platform. Annotate real review data through expert guidance and machine learning methods, by using PU learning algorithms for classification. The example results show that the method in this thesis has good performance of false comment recognition, which provides a new method for consumers and regulatory authorities, and has practical application value.