SetConv: A New Approach for Learning from Imbalanced Data

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1284–1294, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1284 SetConv: A New Approach for Learning from Imbalanced Data Yang Gao1, Yi-Fan Li1, Yu Lin1, Charu Aggarwal2, Latifur Khan1 1The University of Texas at Dallas, USA 2IBM T.J. Watson, USA {yxg122530,yli,yxl163430,lkhan}@utdallas.edu,charu@us.ibm.com Abstract For many real-world classification problems, e.g., sentiment classification, most existing machine learning methods are biased towards the majority class when the Imbalance Ratio (IR) is high. To address this problem, we propose a set convolution (SetConv) operation and an episodic training strategy to extract a single representative for each class, so that classifiers can later be trained on a balanced class distribution. We prove that our proposed algorithm is permutation-invariant despite the order of inputs, and experiments on multiple large-scale benchmark text datasets show the superiority of our proposed framework when compared to other SOTA methods. 1 Introduction In many real-world NLP applications, the collected data follow a skewed distribution (Deng et al., 2009; Fern´andez et al., 2013; Yan et al., 2017), i.e., data from a few classes appear much more frequently than those of other classes. For exam- ple, tweets related to incidents such as shooting or fire are usually rarer compared to those about sports or entertainments. These data instances of- ten represent objects of interest as their rareness may carry important and useful knowledge (He and Garcia, 2009; Sun et al., 2007; Chen and Shyu, 2011). However, most learning algorithms tend to inefficiently utilize them due to their disadvantage in the population (Krawczyk, 2016). Hence, learn- ing discriminative models with imbalanced class distribution is an important and challenging task to the machine learning community. Solutions proposed in previous literature can be generally divided into three categories (Krawczyk, 2016): (1) Data-level methods that employ under- sampling or over-sampling technique to balance the class distributions (Barua et al., 2014; Smith et al., 2014; Sobhani et al., 2014; Zheng et al., 2015). (2) Algorithm-level methods that modify existing learners to alleviate their bias towards the major- ity classes. The most popular branch is the cost- sensitive algorithms, which assign a higher cost on misclassifying the minority class instances. (D´ıaz- Vico et al., 2018). (3) Ensemble-based methods that combine advantages of data-level and algorithm- level methods by merging data-level solutions with classifier ensembles, resulting in robust and ef- ficient learners (Galar et al., 2012; Wang et al., 2015). Despite the success of these approaches on many applications, some of their drawbacks have been observed. Resampling-based methods need to ei- ther remove lots of samples from the majority class or introduce a large amount of synthetic samples to the minority class, which may respectively lose im- portant information or significantly increase the ad- verse correlation among samples (Wu et al., 2017). It is difficult to set the actual cost value in cost- sensitive approaches and they are often not given by expert before hand (Krawczyk, 2016). Also, how to guarantee and utilize the diversity of clas- sification …

Content truncated to 500 words