Improving the Performance of Sentiment Classification on Imbalanced Datasets With Transfer Learning

Abstract

In recent years, many sentiments classification models, such as deep learning models and traditional machine learning models, claim that they can achieve state-of-the-art performance in sentiment analysis problems. Admittedly, this is based on the premise that the training samples are class balanced. However, in the real world, the training data sets we can get are often imbalanced, which will cause the trained classifier to tend to predict the test samples into a majority, making the recall of minority very low. In order to minimize the influence of the imbalanced data class on the model performance, a transfer learning method based on a convolution neural network is proposed in this paper. First, we use a CNN-based model for pre-training in the class-balanced source domain data set, before transferring the model to the target domain for fine-tuning to improve the recall of minority class; furthermore, we propose a transfer learning-based under-sampling technique, which can under-sample the majority class in the target domain. In the data set after under-sampling, we again fine-tune the pre-trained model, so that the recall and precision of the minority class have been greatly improved. The experiments on real-world data sets show that our proposed under-sampling method has obvious advantages compared with others.