Effect of Sampling Strategies on Fine-grained Emotion Classification in Microblog Text

This study investigates the effect of diverse training samples on machine learning model performance for fine-grained emotion classification. Using four different sampling strategies (random sampling, sampling by topic and two variations of sampling by user), we found the class distribution of 28 emotion categories to differ across the samples produced by each sampling strategy. However, combining different sampling strategies is complementary in generating sufficiently diverse training examples for the emotion classifiers. Based on support vector machine (SVM) and Bayesian network learning algorithms, our findings show that a classifier trained on combined data from the four sampling strategies performs better and is more generalizable than a classifier trained only on data from a single sampling strategy. Demonstrating how the diversity of the training samples affect the performance of emotion classifiers is the main contribution of this study.