برچسب‌گذاري نظرات فارسي با استفاده از يادگيري فعال و نيمه‌نظارت شده جهت پيش‌بيني ريزش مشتريان

Labeling Persian Reviews using Active an‎d Semi-supervised Learning to Customer Churn Prediction

مهندسي فناوري اطلاعات

Nowadays, predicting customer churn in all businesses is considered an important an‎d necessary task so that organizations an‎d companies can identify at-risk customers in a timely manner an‎d prevent their departure. While numerous efforts have been made in the field of customer churn prediction, most of these studies have been conducted using structured data an‎d mainly for the telecommunications industry, with limited use of textual data. Additionally, in the Persian language domain, very few studies have been conducted on customer churn prediction using textual data an‎d in these studies, the labeling of reviews has been entirely performed manually. Given the current lack of extensive labeled textual data for customer churn prediction an‎d the fact that Persian textual data is mainly labeled for sentiment analysis an‎d customer satisfaction prediction, this research aims to address this gap by employing a method that minimizes human effort to label a large volume of textual data for customer churn prediction. To achieve this, it was decided to label Persian reviews using active learning an‎d semi-supervised learning. The proposed method of this study is as follows: initially, a labeled dataset is created with the assistance of three experts, an‎d a deep learning-based model is trained on this dataset. Subsequently, using active learning, a number of unlabeled reviews from the Digikala website are selec‎ted an‎d labeled by three experts with the help of explainable artificial intelligence. Finally, the remaining unlabeled reviews obtained from the Digikala website are labeled using semi-supervised learning. It is worth mentioning that for semi-supervised learning, three methods are used: "self-training," "label propagation," an‎d a "hybrid" method that combines these two approaches. The results indicate that using active learning an‎d semi-supervised learning ("self-training" method) is an efficient approach for labeling Persian reviews for customer churn prediction. This approach not only minimizes human effort an‎d reduces costs but also improves eva‎luation metrics by several percentage points an‎d enhances model performance compared to before implementing this method. It is noteworthy that the findings of this research are applicable to any business, as this study was not conducted for a specific bran‎d, an‎d all products of online stores were included in the research.

6