Classification of Advertisement Text on Facebook Using Synthetic Minority Over-Sampling Technique

Understanding in consumer behavior is an important task in the field of marketing. Dentsu's AISAS model is a model that has been proposed to describe consumer behavior. The model defines reaction when the consumer has seen advertising into five stages: attention, interest, search, action, and share. In this paper, advertisement text datasets from Facebook were labelled as the stages of AISAS model and learned to be classified by machine learning algorithms. Nevertheless, like many other real-world data, our dataset had imbalanced class distribution. The classifier algorithms tend to predict mostly the majority class. To overcome this problem, synthetic minority over-sampling technique (SMOTE) was adopted and also combined with chi-square based feature selection technique. Varieties of feature sizes based on various classifier algorithms were compared. In the appropriate feature size, SMOTE could improve the classification performance in terms of recall and F1 score.