Estimating Evaluation of Cosmetics Reviews with Machine Learning Methods

This paper presents methods for estimating evaluation of cosmetics reviews, i.e., for assigning scores to cosmetics reviews, by using three kinds of machine learning methods: support vector machine (SVM), stacked denoising autoencoder (SdA), and convolutional neural network (CNN). The experimental results show that (1) using words with various parts of speech (POSs), not only nouns, as features in vectorizing review text is effective, (2) selecting features on the basis of dependency analysis is effective, (3) the three machine learning methods have almost the same estimation precision and are much higher than a rule-based baseline method, and (4) the training cost of SVM is extremely lower than the other two methods, and SVM therefore performs the best among the three methods.